Changelog

​

2026-10-13

​

Gemini 3.8 Flash now honors reasoning_effort on chat completions, and its thinking shows in usage

​

2026-10-09

​

Generate images with GPT Image 2.5 Flare, which keeps a character consistent across reference images

• Keep one character across a series — pass a few reference photos and get the same face and details in new scenes
• Render exact text — wordmarks, slide titles and social cards spelled as written
• Edit from a reference — change setting, lighting or outfit while the subject stays the same

​

Generate cinematic video clips with Kling 3 Pro, from a prompt or from a first-frame image

• Hero shots — photoreal people and smooth camera motion for brand video
• Animate a still — pass a product photo or portrait as image_url and the prompt drives the motion
• Predictable price — only cfg_scale is read from extra, so the clip you get is the one the price was held for

​

Turn one portrait and one voice or song track into a lip-synced talking or singing clip

• Talking heads from a photo — a presenter or narrator speaking your recording
• Lip-synced singing — a character singing your own track, real vocal kept
• Any language — the mouth follows the audio itself, not a transcript

​

Generate a full song with vocals from a prompt and your own lyrics, in Vietnamese or any other language, with Lyria 3.5

• Sing your own lyrics — Vietnamese lyrics come back sung in order
• One price per song — no per-second meter, so a longer take costs the same
• Not charged when no song comes back — a prompt the safety filter refuses costs nothing

​

Generate video again, including clips with native speech and sound, with Veo 3.1 Fast

• Clips that talk — dialogue, ambient sound and lip-sync from one prompt
• Start from your own frame — pass image_url to animate a still
• Vertical or wide — aspect_ratio 16:9 or 9:16

​

A failed image, video or audio job on the Logs page now shows the same error code as GET on the job

​

2026-10-08

​

Three media model ids that can no longer be served are retired, so no page, alias or example offers a model that cannot answer

​

2026-10-07

​

/v1/messages honours the Anthropic thinking parameter and returns the model’s reasoning as thinking content blocks, streamed and not

​

2026-10-06

​

Your unverified accounts and API keys are now strictly protected from unauthorized changes.

​

You will no longer accidentally accumulate debt when running embeddings or reranking with an empty balance.

​

2026-10-04

​

12 prices fall and 7 rise — every one computed from what the request actually costs to serve

​

The default API key can now be disabled and deleted like any other, and the key list comes back newest first instead of in whatever order Postgres felt like

​

A kyma-gen run that takes longer than the sync wait now keeps going to its result instead of stopping with a timeout

​

A refusal is never answered by another creator’s model under the name you asked for

​

When Kyma falls back to another model, the bill follows that model’s price and never exceeds the price of the model you asked for

​

2026-09-29

​

Creating, disabling and deleting API keys now requires a dashboard session, so a leaked API key cannot mint new keys that outlive it

​

2026-09-28

​

Audio understanding is billed from the audio itself, so the price no longer depends on a duration you type

​

2026-09-24

​

Every API error now carries a stable string code, a category, whether to retry, a link to its docs entry and a request id you can quote to support, and /v1/messages answers every error in Anthropic’s own format

​

A stream that fails after it has started now ends with a clear, retryable error instead of passing through the upstream’s own message, and /v1/messages clients see a proper error event

​

2026-09-22

​

Billing settings and admin now require a dashboard session, so an API key used for inference cannot change auto top-up, saved cards or billing details

​

Tier rows in GET /v1/auth/limits report audio concurrency per capability, the same shape GET /v1/limits/tiers already uses, and four error answers now tell you what to do instead of carrying internal routing details

​

Request logs show which key made each call by its name and last six characters; logs no longer return full key strings

​

cancelling tells you plainly when a job can no longer be stopped, instead of reporting it cancelled

​

Sending an embedding or reranking model to /v1/messages returns a clear 400 naming the right endpoint instead of a chat answer

​

GET /v1/auth/limits now shows how many requests and tokens you have used this minute, ?tools=true on GET /v1/models lists only models that can call tools, and chat responses carry usage.cached_tokens on every call

• Know your headroom before you hit a 429 — GET /v1/auth/limits returns this minute’s rpm_used and tpm_used, with rpm_remaining and tpm_remaining as the limit minus them and resets_at at the start of the next minute. Checking your limits does not use up a request.
• Filter the catalogue by what a model can do — GET /v1/models?tools=true and ?structured_outputs=true return exactly the models whose supports_tools / supports_structured_outputs is true.
• Read cache hits in one place — every /v1/chat/completions response, streamed or not, carries usage.cached_tokens: the cached input tokens the request was priced with, 0 when nothing was cached. prompt_tokens_details is still passed through as the model sent it. No price changed.

​

2026-09-21

​

Grok 4.7, xAI’s newest flagship for coding, agentic tasks and knowledge work, is callable as grok-4.7 with a 500K context window and image and file input

• Move a Grok agent to the newest release: change the model id from grok-4.6 to grok-4.7
• Run a long coding task in one call: send up to 500,000 tokens of code, images or files in one request
• Pin an older release while you compare: keep grok-4.6 in production and route a test slice to grok-4.7

​

An image or video job now shows the same cost on every endpoint that returns it, and says how much of the hold came back

​

The catalogue, pricing, docs and llms.txt stop offering two models Google is shutting down, so nobody starts new work on them, while existing calls keep working until each retirement date

​

2026-09-20

​

When every provider for a model is failing, you get 503 with Retry-After instead of a 500 with no guidance, and the failure shows up in your request log

​

Every kyma-gen generation now tells you what it held, what it charged and what it gave back, including when you cancel it

​

POST /v1/messages now tells you what a non-streamed call cost, and browser code can finally read Kyma’s response headers

​

A request the fallback cannot serve now gets a clear 400 telling you to call the primary endpoint, instead of an unexplained 500

• Know why the call failed — the answer names the reason and the fix (use /v1/chat/completions for :online models, or drop :online) instead of a bare server error.
• Stop retrying something that will never work — a 400 tells an SDK not to retry; a 500 invites it to try the same request again.
• Tell a real outage from a request the fallback does not support — a 500 used to mean both.

​

Kyma stops advertising video models it no longer serves — the description you read now lists exactly what the catalogue can do

​

A public Muse share link no longer carries the account id of whoever shared it, and neither do the URLs of files you upload

files were stored under an account-scoped path

files are stored under the job or upload id alone

​

A streamed turn that answers with a tool call or with reasoning is now charged and logged like any other, instead of silently costing nothing and leaving no record

​

The welcome email, the repo’s front page and the fallback backend’s own description stop naming a kind of model nobody can buy

​

2026-09-19

​

A model that answers a question with a choice and a confidence instead of prose, at POST /v1/decisions

• Route a request — ask which tier should serve it, act on the answer above a confidence you pick, and send the rest to a full model.
• Gate an action — ask whether a command or a tool call should run, and escalate the uncertain ones to a human or a larger model.
• Rate or label — sort tickets, messages or documents into buckets you define in the request itself, with no fine-tuning.

none — the path is new to customers

POST /v1/decisions

​

Muse Spark 1.3, Meta’s release after 1.2, is callable as muse-spark-1.3 with a 1M context window and multimodal input

• Move a Muse Spark agent to the newest release — change the model id from muse-spark-1.2 to muse-spark-1.3 and keep the same price
• Reason over a long file set in one call — send up to 1,048,576 tokens of text, images, video, files or audio in one request
• Pin an older release while you compare — keep muse-spark-1.2 or muse-spark-1.1 in production and route a test slice to muse-spark-1.3

​

Transcription checks your balance against the file you sent, before the audio leaves Kyma, and a transcription Kyma has already paid for is always charged instead of being refused afterwards

​

2026-09-13

​

A text document sent to /v1/messages now reaches the model, so the answer is about the document you sent instead of one the model never saw

​

Images sent to /v1/messages now reach vision models, so the answer describes the picture you sent instead of one the model never saw

​

A streamed /v1/messages response now ends with exactly one message_delta and one message_stop, and that message_delta reports input_tokens and output_tokens, so clients that track usage from the stream see the real numbers

​

tool_choice none and disable_parallel_tool_use now reach the model, so a request that says no tool calls, or one tool call at a time, gets that

​

during a promotion both endpoints now quote what the model costs today, with the list price beside it, instead of the list price alone

​

A request the model rejects as invalid now returns that status (400, 413 or 422) instead of 503, so SDKs stop retrying a request that cannot succeed and the error says to change the request

​

turning parallel tool calls off now means at most one tool call per turn on every model, including models that ignored the setting

​

an oversized request is refused at once with a 413 that names its size, before Kyma reads the body, on every chat endpoint

​

2026-09-12

​

Two new models — DeepSeek V4.1 Flash, the lab’s successor to both V4 Flash and V4 Pro with a 1M context window and image input at the Flash price band, and GPT-6 Astra, OpenAI’s GPT-6 flagship with a 1.05M context window. deepseek-v4-pro retires on 2026-09-14 in favour of deepseek-v4.1-flash, and from now on a model with an announced retirement date returns the retired-model error naming its replacement from that date

• Pin a DeepSeek id that will still be there next month — point an agent at deepseek-v4.1-flash instead of deepseek-v4-pro, which the lab reroutes on 2026-09-14
• Read a screenshot on a Flash budget — send an image to deepseek-v4.1-flash and get a transcript or answer at $0.27 /$0.81 instead of a flagship rate
• Run a hard reasoning task on OpenAI’s newest flagship — call gpt-6-astra with a 1.05M-token context and prompt caching, and read the exact cost in usage.cost
• Move off deepseek-v4-pro before 2026-09-14 — switch the id to deepseek-v4.1-flash; on the day the old id returns the retired-model error naming the replacement instead of silently answering with a different model

​

The catalogue now lists only the media models Kyma actually carries, so a model you can pick is a model that works; paused SKUs say so up front instead of failing at the provider

• Keep generating images — gpt-image-2 for text-in-image, flux-2-pro for photoreal, recraft-v4 and recraft-v4-vector for design and SVG, flux-kontext-pro to edit an image you already have.
• Keep generating speech — swap eleven-v3 for eleven-multilingual-v2, or minimax-speech-turbo when cost per character matters more than expressiveness.
• Keep transcribing — nothing changed: whisper-v3-turbo for timestamps and captions, gpt-transcribe or gemini-3.5-transcribe for accuracy on conversational audio.

​

2026-09-11

​

Chat completions on the Bun runtime answer again instead of 500ing on every request

​

Two model ids that were being answered by a different model are retired with a named replacement, one id is pinned to the build its route actually serves, and two retirement dates are corrected

​

An upload no longer fails on a generic Content-Type or a WebM/OGG/FLAC file; a segment format on a model without timestamps is refused up front instead of returning an empty file that was billed; a silent file is a 422, not a 502; every transcription error names a model that can serve the request

​

A transcription call that omits response_format works again on every model; the default follows the model instead of assuming segment timestamps

​

2026-09-08

​

Saving or updating a chat over 512 KB is refused up front with 413 instead of being stored, and an error while sharing or cancelling a job no longer carries internal database text

​

2026-09-07

​

A comparison you build yourself now becomes a real, listed page. Pick any two to four models on a /compare page and the combination shows up on /compare under “Recently compared”, and once it has been opened on three separate days and every model in it has enough measurement behind it, it goes into the sitemap and can be found in search. Three and four model comparisons are no longer excluded from search on principle. The conclusion sentence on each page (“X is ahead at 99.4%”) now sits above the chart instead of under it, so the answer is the first thing read.

• Compare four models and keep the page — pick any two to four models on a comparison page; the combination gets a URL, appears on /compare under “Recently compared”, and can be shared like any other page.
• Find what other people compare — /compare now ranks the combinations opened in the last 30 days by how many times each has been opened, and GET /v1/compare/popular returns the same list to an agent with no authentication.
• Read the answer first — each comparison states its conclusion in a sentence above the chart, so an answer engine quoting the page quotes the finding rather than a caption.

none — /compare and /compare/<slug> keep their URLs, and a non-canonical ordering still permanently redirects to the sorted one

POST /v1/compare/view and GET /v1/compare/popular, both public and unauthenticated

​

Every model on GET /v1/models now says whether an account that has never added credits can call it, so you can tell before you send a request instead of finding out from a 402

• Filter before you send — curl "https://kymaapi.com/v1/models" | jq '[.data[] | select(.free_tier)] | map(.id)' lists everything a brand-new key can call, so a first request never spends a round trip on a refusal
• Ship an onboarding that works on day one — point the free plan of your app at the free-tier ids and let paying users unlock the rest, instead of hard-coding a list that goes stale the next time a price moves
• Let an agent pick its own model — in the Kyma MCP server, recommend_model now recommends only free-tier models to a key that has never been topped up and says why, so a coding agent configures itself with a model it can actually call

​

Signing in now also sets a session cookie, so the dashboard can render your sidebar and email on the server instead of waiting for JavaScript to read localStorage and ask who you are

• Stay signed in across a hard refresh without a flash — the browser presents the cookie on the first request, so the page can know who you are before any JavaScript runs
• Sign out everywhere it counts — POST /v1/auth/logout and DELETE /v1/auth/session/cookie both clear the cookie, so a shared machine does not keep a signed-in shell
• Keep using Bearer tokens exactly as before — curl, the SDKs and every agent send Authorization and are never asked for a CSRF header

Set-Cookie: __Host-kyma_session=ks-…; Path=/; Max-Age=…; Secure; HttpOnly; SameSite=Lax
Set-Cookie: kyma_csrf=…;              Path=/; Max-Age=…; Secure; SameSite=Lax

none — no path moved

POST/DELETE /v1/auth/session/cookie, GET /v1/auth/session/expiry

​

Your per-model usage breakdown answers in a fraction of a second instead of seconds when you do not name a time window, because it now defaults to the last 30 days rather than scanning your whole account history

​

Every place that mentions the $0.50 signup credit now says which models it buys — the free tier — instead of implying it covers the whole catalogue, so you find out before a request that a video, image or frontier model needs a top-up

none — no path moved

https://kymaapi.com/pricing#free-tier — a new anchor listing the free tier and its media models

​

Every page now shows the price you are actually charged, to the digit

• One price, everywhere — /pricing, /models, each model page, /rankings and the machine-readable pricing.md and models.md all print the published rate to the digit
• /pricing looks like /models — each row now carries the creator’s mark, the context window, and the model’s measured 30-day uptime, and links to its page
• “n/a”, not a dash — 45 of the 82 token models publish no cached-input rate. The cell now says which of the two reasons applies: the model does not support prompt caching, or it does and bills repeats at the full input rate
• No hand-assigned scores — the rankings cost tab ranked models by a “quality” number typed into the source for 26 of them and defaulted to 70 for the rest. It is gone; the tab sorts on measured cost

​

The rankings Performance board stops going blank when one leaderboard query fails

• See the board — a leaderboard query that fails is retried on the next request instead of being pinned for five minutes
• Know when it is incomplete — a partial response carries X-Kyma-Partial: <query names> and Cache-Control: max-age=30, a complete one carries max-age=300
• Find it in the logs — every failed sub-query is logged with its name and the database’s error message, which none of them were before

​

The rankings page stops showing a blank Performance board for five minutes after the API has recovered

• See a board that came back — an incomplete answer is re-read after 30 s, and the page that renders it expires with it
• Same for an outage — if the API cannot be reached at all, the empty state is no longer pinned for five minutes either
• Nothing slower when it works — a complete board is still read once and cached the full five minutes

​

The rankings board ranks on reliability, prices every model in the unit it is sold by, and reports the same uptime as the rest of the site

• Scan for what works — reliability is the first column and the default sort, with volume breaking ties
• Compare like with like — embeddings and audio have their own table below the board, showing only what is defined for them
• Read one uptime — the Usage tab reads the same measurement as every other Kyma surface, and says “no data yet” for the 31 of 117 ranked models with too few observations in the window instead of guessing
• Trust the prices — models sold by the minute or by the character are priced in that unit. Four audio models were showing a per-token default of $0.200 /$0.600 that was nobody’s real rate
• Read only measurements — the Value column is gone. It folded a hand-assigned quality score into speed, reliability and cost, so a column of Kyma measurements carried one number Kyma never measured

​

2026-09-06

​

The dashboard shell loads a 190-byte identity instead of a 14 KB key dump on every page, per-key spend lives on the keys endpoint, and the Usage and Logs pages answer in well under a second

• Admin pages accept the signed-in admin session; no separate admin key is required.
• The Usage page’s daily aggregate and the Logs page’s model list are indexed and computed in the database; both answer in well under a second.
• Each dashboard page reports when its data is on screen (dashboard_data_ready), so slow loads are measured by what you waited for, not by when the heading painted.

​

The MCP sign-in screen now has Continue with Google, so accounts created with Google can connect Claude, Cursor, Codex and ChatGPT without a password

​

2026-09-05

​

qwen-3.6-plus gets cheaper — the published rate follows what it actually costs to serve

​

The Usage page loads in well under a second, and 7d / 30d / 90d show the range they say

• Daily buckets follow your browser’s timezone (they were UTC).
• By-model cost is exact on every range.
• While a new range loads, the tiles and chart dim instead of showing stale numbers.

​

2026-09-04

​

When the measurement query behind model uptime is unavailable, the endpoint now says so instead of reporting every model as unmeasured, and any unexpected failure at the edge comes back as readable JSON a browser can actually display

​

Saved chats, one-click unsubscribe, sharing a generated image or video, and cancelling a kyma-gen job are now answered at the Cloudflare edge instead of being relayed to the origin

​

Rankings, model uptime, cost per session, model recommendations, public stats and badges, the media registry, generation pricing and public Muse shares are now answered at the Cloudflare edge instead of being relayed to the origin

​

2026-09-03

​

A request for a retired Imagen 4 model now names its replacement instead of failing upstream with nothing to act on

POST /v1/images/generations with model imagen-4, imagen-4-fast or imagen-4-ultra

POST /v1/images/generations with model nano-banana-3-flash

​

Gemini 3.8 Flash and Claude Fable 5.1 are available by name on the chat endpoints

• Agentic coding at Flash cost — point a coding agent at gemini-3.8-flash for multi-step tool use over a large repository; 1M context and 64K output mean the whole task fits one request.
• Long multimodal review — send video, audio, images and files to gemini-3.8-flash in one call and get a long written analysis back.
• The hardest reasoning and long-running agents — use claude-fable-5.1 where Fable 5 was the pick: long-horizon agentic workflows and agentic coding, with 1M context and 128K output.

​

Two open-weight multimodal models are available by name — Qwen 3.8 27B and DeepSeek V4 Flash Vision (experimental)

• Read screenshots and diagrams on a budget — send an image with the question to deepseek-v4-flash-vision-exp at the V4 Flash price tier.
• Video and document understanding on open weights — qwen3.8-27b takes video and image input at 1M context, with tool calling, on an Apache-2.0 checkpoint you can also run yourself.
• An open alternative to the closed Qwen 3.8 tiers — same family, same API shape, open weights.

​

Promotional prices on five models now show as promotions with an end date, and GPT-5.6 Sol is listed at the price you were already being billed

​

qwen-3.7-max is listed at the price of the route that now serves it, so the page and the bill agree again

​

The 43 image, video, speech, music and transcription model pages now show the price in the unit the model bills in, and only the features it has

​

August 14, 2026

​

Gemini 3.1 Pro is live at $2.70/$16.20 per million tokens — the first Gemini Pro tier on Kyma — and Nano Banana 3 Flash is repaired after its upstream preview endpoint stopped answering

• Reach for a Pro tier without leaving Gemini — hard reasoning and long-horizon agentic work at 1M context, with the same OpenAI-compatible call and the same automatic failover as everything else
• Know the long-prompt rate up front — the 200K step is published next to the base rate on /pricing and in /v1/models, not discovered on the invoice
• Nano Banana 3 Flash answers again — the upstream endpoint swap was transparent: same nano-banana-3-flash model id, same $0.061 per image

​

Gemini 3.7 Flash is live at $1.01/$5.06 per million tokens — and half that, $0.51/$2.53, while the launch promotion runs

• Stop chunking long replies — a full document translation or a 40,000-token structured extraction now fits in one response instead of a loop that stitches pieces together
• Send video, audio and files to one model — the widest input set in the Flash family, with tool calling, reasoning and structured outputs on the same request
• Check what you actually pay before you commit — /v1/models publishes the promotional rate, the list rate it returns to, and the date it ends, so a budget built on it does not go stale silently

​

7 prices fall and 14 rise — every one computed from what the request actually costs to serve

​

/v1/models now quotes the price you actually pay today, so a model on promotion no longer reads as twice its real cost

​

2026-08-13

​

Claude Fable 5 is live at $13.50 in and$67.50 out per million tokens

• Hard reasoning on one id — 1M context and extended reasoning without leaving the catalog
• Vision and files on the same call — image and file input without a second SKU
• Read the price before you commit — /v1/models publishes the rate in force

​

Grok 4.6 is live at $2.70 in and$8.10 out per million tokens

• Long-horizon agents on one id — 500K context and tool calling without stitching a second model
• Know the long-prompt rate up front — the 200K step is published on /v1/models, not discovered on the invoice
• Vision and files on the same call — image and file input without switching SKUs

​

Qwen 3.8 Max is live at $2.2275 in and$6.684 out per million tokens

• Reason over a long document — 1M context without chunking the prompt by hand
• Send an image with the question — vision on the Max tier, same model id
• Read the cache rate before you commit — /v1/models publishes the cached-input price next to the list rate

​

Qwen 3.8 Max is listed once, at the $2.2275 /$6.684 rate

​

August 3, 2026

​

You are now billed at the exact rates published by the suppliers, eliminating incorrect pricing on several popular models.

​

The model uptime dashboard now accurately reflects supplier availability instead of showing false outages caused by our own probe limits.

​

August 2, 2026

​

1 model got cheaper to call, effective immediately

​

August 1, 2026

​

You can list everything one lab makes in a single request

• Compare two labs — ?creator=anthropic,google puts them side by side before you commit to one
• Audit a family — see every model one vendor makes, and what each costs, in a single call
• Build a vendor picker — group your model dropdown by lab without hard-coding the list

​

One click unsubscribes you from Kyma email, and it sticks

​

You can add live web results to 25 models by appending one suffix to the model id

• Answer from today — ask about a release that shipped this morning and get the release, not the training cutoff
• Keep your model — add :online to the model you already tuned prompts for instead of switching to a search-native one
• Turn it off per request — drop the suffix when the answer should come from the model alone, with no extra call

​

Grok, Meta’s newest line and Claude are all callable with the same key as everything else

• Try a frontier model without a new account — Grok and Claude bill against the key you already have
• Swap in one line — change the model id; the request shape, the fallbacks and the receipts stay the same
• Compare on your own traffic — run the same prompt across labs and read the cost back from usage

​

Published prices follow the cheapest route that can serve you, not whichever one happened to carry the last request

​

What a model says it can do now matches what its provider says it can do

​

Searching the models page no longer throws you somewhere else on the page

​

July 31, 2026

​

You can compare models on measured speed and uptime instead of taking a claim on trust

• Pick on evidence — sort by throughput or uptime instead of trusting a launch post
• Rule a model out — a dash means it has not been measured enough, and says so rather than guessing
• Check before you migrate — see how a model has actually behaved over 30 days, not on the day it launched

​

When a supplier cuts its price, yours falls too — automatically, and within a day

​

Your account is harder to attack, and a database read no longer yields anything replayable

​

The cost in your streaming response is what you were charged, not what your request cost Kyma

​

July 30, 2026

​

Older clients that speak the legacy completions shape work without changing your code

• Keep an old client — point it at Kyma without rewriting it for the chat shape
• Migrate on your own schedule — move to /v1/chat/completions when it suits you, not because the API forced it
• Get the same guarantees — an adapter over the same routing, so a legacy call still gets never-die fallbacks

​

You can build retrieval on Kyma without a second vendor for the embedding half

• Index a corpus — embed your documents with the same key that answers questions about them
• Drop a vendor — retrieval and generation stop needing two accounts, two bills and two outage pages
• Pay for input only — there is no output side on an embedding, and you are not charged for one

​

Every model has its own page, with the whole price and a plain-markdown mirror an agent can read

• Read the whole price — input, output and cached, not a rounded headline
• Feed it to an agent — every page has a .md twin at the same path, so a crawler gets facts, not layout
• Check before you call — context window, capabilities and what it is good for, in one place

​

You are billed for the route that actually served you, and never above the published price

​

Pages load faster, and the site stops rendering dark on a light theme

​

July 29, 2026

​

You have until 20 October to move off nano-banana, and Kyma stopped recommending it today

​

Six models joined the catalogue, including the cheapest one Kyma sells

• Kimi K3 — Moonshot’s 2.8T open-weight reasoning model, 1M context, vision in
• Gemini 3.6 Flash and Gemini 3.5 Flash Lite
• Qwen 3.7 Flash — the cheapest model in the catalogue
• GPT-5.6 Terra, priced on list rather than on a promotion that expires
• Gemini 3.1 Flash Live, the successor to the 2.5 voice SKU

• Reach for 1M context — Kimi K3 and Gemini 3.6 Flash both take a repository-sized conversation
• Cut the floor price — Qwen 3.7 Flash is the cheapest model in the catalogue
• Move a voice workload — Gemini 3.1 Flash Live succeeds the 2.5 voice SKU that is retiring

​

Five models were repriced against what they actually cost to serve — two of them had been sold below that

​

Multi-turn reasoning with Gemini 3 keeps its train of thought across turns

​

July 28, 2026

​

A long conversation is never rescued by a model too small to hold it

​

July 27, 2026

​

You can ask the catalogue for exactly the models you need instead of reading all of them

​

An error now tells you whose problem it is, so you know whether retrying will help

​

June 2026

​

Every model got a page with live numbers instead of a specification copied from a launch post

• Check before you switch — read a model’s real numbers rather than its announcement
• Follow a name from the rankings — the table rows lead somewhere now
• Compare two pages — the same fields, in the same order, for every model in the catalogue

​

Six models joined, including two flagship coding models and the strongest open-weight tier Kyma had carried

• Put a coding model behind an agent — GLM-5.2 and Kimi K2.7-Code were added for exactly that shape of work
• Reach 550B open weights — nemotron-3-ultra is the largest open-weight model in the catalogue
• Translate in a live call — the Live Translate SKU runs on the realtime path, not as a batch job

​

Speech and transcription stopped failing outright when one supplier did

​

June 19, 2026

​

2 new flagship models: GLM 5.2 and Kimi K2.7 Code

• glm-5.2 — Zhipu’s newest frontier-open flagship. 744B MoE, 1M-token context (up from ~200K on GLM 5.1), #1 open-weight on the Intelligence Index v4.1. $1.89/$5.94 per 1M tokens. The glm-flagship alias now resolves here.
• kimi-k2.7-code — Moonshot’s coding specialist. +21.8% on Kimi Code Bench v2 vs K2.6 with ~30% fewer reasoning tokens, 262K context, text+image input, always-on thinking. $1.28/$5.40 per 1M tokens.

​

June 10, 2026

​

4 new models: Qwen 3.7 Plus, MiniMax M3, Nemotron 3 Ultra, Step 3.7 Flash

• qwen-3.7-plus — successor to qwen-3.6-plus. 1M context, adds vision, $0.54/$2.16 per 1M tokens
• minimax-m3 — agentic coding focus, 1M context, text+image+video input, $0.405/$1.62 (same price as M2.5)
• nemotron-3-ultra-550b — largest US open-weight model (550B MoE), 1M context, $0.675/$3.375
• step-3.7-flash — cheap multimodal flash tier, 256K context, tool calling, $0.27/$1.553

• Reliability: per-attempt upstream timeout — a hung provider now fails over instead of hanging your request
• Billing hardening: settle idempotency, usage estimation fallback, and time-to-first-byte now measured on every request
• deepseek-r1 input price adjusted to $0.74/1M (output unchanged)

​

June 4, 2026

​

ElevenLabs v3 — most expressive TTS + low-latency streaming

• Model eleven-v3 — the most expressive text-to-speech model on Kyma. Audio tags and emotional range for character voices and dialogue, lifelike delivery across 70+ languages. Same $0.405/1K char as Multilingual v2. Call it on POST /v1/audio/speech with any ElevenLabs voice.

• Feature Opt-in low-latency TTS streaming — pass "stream": true on POST /v1/audio/speech to receive audio progressively as it’s synthesized. Time-to-first-audio drops to ~0.4s (from ~1.8s). The response is still a single complete audio/mpeg stream, so existing clients keep working unchanged. Currently applies to the MiniMax speech models.

​

May 2026

​

Speech, music and sound effects arrived, so an audio app no longer needs a second vendor and a second bill

• Narrate something long — send a chapter to /v1/audio/speech and get a file back, billed per character
• Score a video — /v1/audio/music writes to a prompt and bills per second, so a 30-second bed costs what 30 seconds costs
• Clone a voice once and reuse it — upload a sample, get an id, and pass that id anywhere a voice is accepted

​

Image and video generation grew a real catalogue, and pricing that matches how each one actually charges

• Ship vector art — Recraft v4 Vector returns SVG, not a raster you have to trace
• Print at 4MP — the Pro tier exists for output that has to survive being printed
• Generate video by the second — Hailuo 02 at 512p, 768p or 1080p, billed for what it produced

​

Web-grounded answers became callable models rather than a separate search product to integrate

• Ask about this week — a question whose answer changed yesterday gets yesterday’s answer
• Keep one client — same endpoint, same key, same fallback behaviour as the rest of the catalogue
• See the search fee — the per-request charge appears as itself, not hidden in a token price

​

Signing up got harder to abuse and easier for real users on shared connections

​

May 17, 2026 (later)

​

Google media models — 7 new SKUs + public pricing catalog

• Model imagen-4-fast — Google Imagen 4 fast tier, $0.027/image
• Model imagen-4 — Google Imagen 4 default, $0.054/image (recommended)
• Model imagen-4-ultra — Google Imagen 4 print-ready, $0.081/image
• Model nano-banana — Gemini 2.5 Flash Image with native edit-mode (image-in + prompt → image-out), $0.046/image
• Model nano-banana-3-flash — Gemini 3.1 image preview, $0.046/image

• Model veo-3-fast — Google Veo 3 fast tier, 720p no-audio, $0.135/sec (recommended budget)
• Model veo-3 — Google Veo 3 flagship, 1080p with native audio (dialogue + ambient + lip-sync), $0.540/sec

• API GET /v1/pricing — full catalog: text + image + video + audio in one round-trip. Replaces the partial /v1/credits/pricing (kept for backward compat). Cache-Control 60s.
• API GET /v1/limits/tiers — full 5-tier matrix, public, cache-Control 300s. Drives the Rate Limits guide and dashboard signup flow.

​

May 17, 2026

​

Audio infrastructure refresh

• Feature Realtime audio WebSocket proxy — scales to 5000 concurrent sessions (was 50). See Realtime Audio.
• Feature STT URL mode — POST /v1/audio/transcriptions accepts JSON {"audio_url": "https://..."} up to 100 MB. Multipart upload remains capped at 25 MB. See Audio Transcriptions.
• Feature STT never-die failover — when the primary whisper-v3-turbo hits a transient hiccup, Kyma transparently retries, then routes to a timestamp-preserving secondary (whisper-1), and for plain-text transcripts a final tertiary (gemini-3-flash-audio) — same request, same price. Response carries X-Kyma-Fallback (serving model) and X-Kyma-Fallback-Layer when it fires. See Audio Transcriptions.
• Feature Tier override for heavy users — partners and enterprises can request Tier 4 limits without the $1000 lifetime deposit. See Rate Limits — Need higher limits?.

• Update Audio rate limits split into per-capability sub-pools (transcription, understanding, speech). Saturating one no longer blocks the others. See Rate Limits — Audio limits.
• Update Audio concurrency caps raised across all tiers — Tier 4 now 100 total audio slots (was lower).
• Update minimax-music-pro backed by music-2.6 (was music-2.5). API contract and pricing unchanged.

• Ops In-process audio load_factor monitor with Telegram alerts when sustained utilization exceeds 0.70 / 0.85 thresholds.

​

May 1, 2026

• [Model] gpt-image-2 — OpenAI’s flagship image model live on Kyma. Near-perfect text-in-image (multilingual: Japanese, Korean, Hindi, Bengali), reasoning-augmented composition, photoreal output. Quality dropdown low | medium | high: $0.014 / $0.081 / $0.297 per 1024² image. 4 sizes supported (1024², 1024×1536, 1536×1024, 2048²). Single SKU, no -pro derivative — keeps OpenAI’s exact model ID.
• [API] Unified picker taxonomy across image, video, and audio composers. Three top-level tiers (Quality / Fast / Cheap) replace the prior ad-hoc grouping; capability sub-axis surfaces SVG (image) and Speech / Music / SFX / STT / Audio-understand (audio). Hailuo 02 1080p moved Quality → Fast; Quality is now reserved for SOTA-class output (Kling 3 Pro, Seedance 2 Pro, gpt-image-2, flux-2-pro, ideogram-v3, recraft-v4-pro). /v1/models exposes new tier and capability fields with ?tier=quality and ?capability=vector filter params. Backward-compat: cost_tier, quality_tier, latency_tier still returned.
• [Pricing] New per-quality pricing mode in IMAGE_COSTS. Hold and settle thread the quality param end-to-end so a quality=high request books the right amount on the hold (no refund-and-rebill drift on finalize).
• [Reliability] Recovery for multimodal jobs after worker death. Hailuo / Image-01 / gpt-image-2 jobs that exceeded their poll budget after the worker died were leaking forever in processing and holding credits hostage; they now refund cleanly within ~12 minutes for video and ~7 minutes for image. Heartbeat ticker (~1.5 s) keeps in-flight OpenAI calls visible to the sweep so multi-minute high-quality requests don’t get prematurely failed.
• [Fix] not_multimodal validator allowlist now reads from a single MULTIMODAL_PROVIDERS set (was a hardcoded chain). Adding a new model source is one line — same regression that bit MiniMax onboarding before.
• [Fix] Audio composer prompt counter NFC-normalizes Vietnamese (and other stacked-diacritic) text before measuring length. Telex-encoded Vietnamese was inflating UTF-16 code units 3× and tripping false prompt_too_long rejections at ~660 visible characters on 2000-char-cap models.
• [Fix] Per-model audio character limits matching upstream provider caps (MiniMax music: prompt ≤ 200 / lyrics ≤ 600; ElevenLabs music prompt ≤ 2000; SFX prompt ≤ 500). Errors now name the SKU and which field overflowed.

​

April 30, 2026

​

MiniMax bundle — 9 new SKUs across audio, image, video

• New model minimax-speech-hd — $0.140/1K char — production multilingual voice, ~2.9× cheaper than eleven-multilingual-v2.
• New model minimax-speech-turbo — $0.090/1K char — lowest-latency voice on Kyma, ~2.2× cheaper than eleven-flash-v2-5.
• New model minimax-music — $0.045/song flat — Music-2.0 family, ~90× cheaper than elevenlabs-music for non-hero tracks.
• New model minimax-music-pro — $0.210/song flat — Music-2.5+ richer arrangements at production fidelity.
• New endpoint POST /v1/audio/voice-clone + model — $2.10/voice flat — clone from 10s-5min reference audio (multipart).
• New endpoint POST /v1/audio/voice-design + model — $4.20/voice flat — generate a voice from a text description, no reference needed.

• New model minimax-image-01 — $0.005/image flat — cheapest image SKU on Kyma, ~11× cheaper than recraft-v4.

• New model hailuo-02-512p — $0.140/clip — cheapest video tier on Kyma, ~4× cheaper than Kling 2.5 Pro at 6s.
• New model hailuo-02-768p — $0.420/clip — mid tier, balanced quality vs cost.
• New model hailuo-02-1080p — $0.780/clip — full HD hero output, less than half the cost of Kling 3 audio at 10s.

​

Sharing a cloned voice_id with another account is rejected

​

Image catalog refresh — 5 new SKUs

• New model recraft-v4 — $0.054 — replaces recraft-v3 as the daily default. #1 on the HuggingFace Text-to-Image Arena, beats Midjourney V8 / DALL-E 3 / FLUX in human preference. Same price as V3.
• New model recraft-v4-pro — $0.338 — V4 quality at 4MP for print-ready / large-scale assets.
• New model recraft-v4-vector — $0.108 — native SVG output with editable paths and layers. The only generation models on the market shipping true vector files.
• New model recraft-v4-vector-pro — $0.405 — V4 vector at 4MP for print-ready logos and large-scale signage.
• New model flux-2-pro — $0.041–$0.101 — BFL’s 32B flagship (3× larger than Flux 1.1). Photoreal, ~60% accurate text-in-image, unified gen+edit. Cheaper than flux-1.1-ultra at 1MP.
• New API param image_urls: string[] — multi-reference blending for FLUX.2 Pro, up to 10 source images merged into a single output.
• New per-megapixel pricing mode — FLUX.2 Pro bills a flat base for the first megapixel plus a per-extra-MP rate, rounded to the nearest whole MP. Hold uses the requested size; finalize uses the actual output dimensions.
• recraft-v3 and flux-1.1-ultra are now marked legacy. Existing API contracts continue to work; new projects should use recraft-v4 and flux-2-pro.

​

April 29, 2026

​

Audio - 2 new endpoints + 2 SKUs

• New endpoint POST /v1/audio/transcriptions - speech-to-text, multipart upload, OpenAI Whisper API compatible
• New endpoint POST /v1/audio/understand - audio scene Q&A (tone, music, SFX, language, emotion), custom Kyma endpoint
• New models: whisper-v3-turbo at $0.0009/min, [`gemini-3-flash-audio`](/models/gemini-3-flash-audio) at$0.000648/min
• Per-minute pricing - both endpoints bill in 1-minute increments, rounded up. 1-hour file: $0.054 transcribe +$0.039 understand = $0.093 total
• New aliases - model: "transcribe" and model: "audio-understand" ride forward when underlying SKUs change
• Audio rows now flow through the same V2 ledger as text - visible on /logs, /rankings, and admin scorecards
• Companion CLI watch-cli - open-source orchestrator that gives any agent eyes and ears for any social video URL (~50x cheaper than full multimodal LLM analysis)

​

April 26, 2026

​

Video Generation - 5 new models

• New endpoint POST /v1/videos/generations - async, returns 202 with a job_id; poll GET /v1/jobs/{id} for the result
• New models: kling-2.5-pro, kling-3-pro, kling-3-pro-audio, seedance-2-pro, seedance-2-fast
• Per-second pricing - $0.0945 to$0.410 per second of video, billed against actual clamped duration
• Per-SKU duration caps - Kling 3 family and Seedance support up to 15s; Kling 2.5 stays at 10s
• Hold-and-finalize billing - failures refund in full; idempotency keys supported end-to-end
• T2V or I2V from a single endpoint - pass image_url to switch any video model into image-to-video mode

​

April 25, 2026

​

DeepSeek V4 — Pro and Flash

• deepseek-v4-pro — 1.6T (49B active) flagship for top reasoning and complex coding. $2.35 /$4.70 per 1M.
• deepseek-v4-flash — 284B (13B active) value tier. Same family behavior at the lowest V4 price. $0.19 /$0.38 per 1M.
• 1M context window, 65K max output, tool calling and structured outputs supported on both.
• deepseek-v3 stays available as the previous-gen stable baseline; older workloads do not need to migrate.

​

Image Generation — Week 1

• New endpoint POST /v1/images/generations — async, returns 202 with a job_id; poll GET /v1/jobs/{id} for the result
• New models: flux-1.1-ultra, flux-kontext-pro (image-edit), ideogram-v3, recraft-v3
• Pay-per-image billing with hold-and-finalize semantics — failures are refunded in full
• Idempotency key supported end-to-end; safe to retry POSTs without double-charging

​

April 23, 2026

​

API Reliability and Platform Changes

• Per-model soft weight + latency tracking (#8)
• Ship Kimi K2.6 + fix kyma-agent auto-update EACCES (#9)
• /models nav link + response sanitization fix (#14)
• Routing — broadened fallback chain for deepseek-v3 (#7)
• Sync Kimi K2.6 across docs, landing pages, integrations (#11)

​

April 21, 2026

​

Product and Dashboard Updates

• AI discoverability + models SEO + billing fixes (#3)
• Public model detail — tabbed code blocks like dashboard
• Tuned deepseek-v3 fallback chain for better latency
• AI agent discoverability + public /models SEO pages (#2)
• Admin_top_users RPC — power user analytics without row limit

​

April 19, 2026

​

Product and Dashboard Updates

• Chat playground — searchable model picker, model ID copy, cost per message
• Billing page redesign — Anthropic-style single input, price breakdown, auto top-up inputs
• Billing page redesign — Anthropic-style layout
• Billing page — card editing, invoice links, tier display, unified layout
• Billing — inline card modal, invoice creation, estimated taxes, setup checkout

​

April 17, 2026

​

Agent, Install, and Runtime Improvements

• Switch tier system from spend-based to deposit-based
• Added a new fallback layer for qwen-3.6-plus
• Correct pricing link to docs.kymaapi.com/pricing
• Bump kyma-agent to 0.1.6 and ter bootstrap to 0.1.1
• Qwen-3.6-plus routing tuned to serve the real closed-weight model

​

April 16, 2026

​

Kyma Agent v0.1.12 — KYMA.md context + MCP servers

• KYMA.md body injection — project rules now prepend to every turn’s system prompt, not just the initial one. Precedence: ~/.kyma/agent/KYMA.md → ./KYMA.md → ./KYMA.local.md, with CLAUDE.md and AGENTS.md as project fallbacks. Edits apply on the next turn — no restart needed.
• /init — scaffold KYMA.md in one step. Detects JS/TS, Python, Rust, Go, Ruby, PHP and their frameworks, then proposes a starter file with frontmatter (model, thinking) + Stack / Conventions / Key files / Agent behavior sections. Preview before write.
• /mcp — Model Context Protocol servers — bring any MCP-compatible tool server into Kyma. Configure in ~/.kyma/agent/mcp.json (user) or ./.kyma/mcp.json (project). Subcommands: /mcp to list, /mcp enable <name>, /mcp disable <name>, /mcp test <name>. Tools register automatically at session start as mcp__<server>__<tool>.
• /status consolidation — merged /doctor, /balance, /usage into a single /status. It now shows account, credits, tier limits, lifetime spend, session totals, API latency, local diagnostics, and MCP health in one view.

​

GLM family from Z.AI

• 3 new models added, bringing the active catalog to 16 models:
  • glm-5.1 — flagship long-running coding agent, repo-scale engineering, 203K context
  • glm-4.5-air — cheap agentic bulk tasks, 131K context
  • glm-4.7-flash — cheap long-context throughput, 203K context
• Implicit caching — 50% off on cache hits where the underlying infrastructure supports it.
• Kyma Agent v0.1.11 — the /models slash command now lists GLM 5.1, GLM 4.5 Air, and GLM 4.7 Flash alongside the existing 9 curated models.
• Auto-failover — every GLM model has multi-layer fallbacks so requests keep flowing even if a backend is unavailable.

​

April 15, 2026

​

Kyma Agent v0.1.8

• Fixed /doctor diagnostics in kyma after the 0.1.7 bridge release.
• kyma-ter remains pinned to 0.1.7; this was a package-only hotfix for the bundled ESM diagnostics path.

​

April 11, 2026

​

Kyma CLI v0.3

• Historical note: this entry describes the April 11 launch state. Kyma Agent and Kyma Ter have since been retired.
• kyma command — At that time, install was described as npm install -g kyma-api, then just type kyma to start an interactive chat session from your terminal. This CLI has since been retired.
• Interactive model picker — Type /model in chat to browse active models with arrow keys. Or /model deepseek-r1 to switch instantly.
• Slash commands — /model, /models, /balance, /clear, /help, /exit — manage your session without leaving the chat.
• Pipe mode — cat error.log | kyma "fix this" or git diff | kyma "review this". Auto-detects non-TTY and outputs clean text.
• Device code login — kyma login opens your browser, auto-fills the code, copies to clipboard. Supports Google OAuth.
• JSON mode — kyma models --json for scripts and CI. Auto-quiet in non-TTY environments.

​

April 10, 2026

​

Higher Limits, Better Emails

• 2.5x higher token limits — Free tier now gets 200K tokens/minute (was 80K) and 30 RPM (was 20). Your coding agents can run longer sessions without hitting walls. Tier 1+ also increased proportionally.
• Transactional emails — You’ll now receive a welcome email on signup, a receipt after every purchase, and a heads-up when your balance is running low. Auto top-up failures also notify you immediately.
• Model expansion — Added MiniMax M2.7, Nemotron 3 Super, Step 3.5 Flash, GLM 4.5 Air, Gemma 4 26B MoE at that time. Model grid reorganized into Recommended / Coding & Agents / Fast & Long Context categories.
• Compare page — New /compare page with honest Kyma vs other LLM gateways and direct APIs comparison, including benchmark data and migration snippets.
• Prompt caching as USP — 48% cache hit rate on heavy users, 23% average cost savings. Now highlighted on homepage.

​

April 8, 2026

​

Models & Pricing

• DeepSeek V3 + R1 — added as primary models with multi-layer fallbacks. DeepSeek V3 is GPT-5 class quality. DeepSeek R1 is a reasoning model 96% cheaper than o1.
• Expanded backend redundancy — additional infrastructure for DeepSeek, Llama, and Qwen models. Adds redundancy and lower latency.
• 4 models disabled — removed a low-yield Qwen 3 235B variant (high failure rate), gpt-oss-20b, llama-3.1-8b, gemma-3-27b (superseded by better models).
• Pricing audit — corrected 8 model prices so every listed price covers full multi-route redundancy. If a backup route is too expensive, we remove it rather than raise your price.
• 19 active models at that time with multi-layer infrastructure redundancy.

​

April 7, 2026

​

⚡ Reliability & Performance

• Auto-failover — if a model’s primary infrastructure is down, requests automatically retry on backup providers. Most failures are invisible to you.
• Faster responses — reduced internal overhead by ~250ms per request through smarter caching.
• Better rate limiting — rate limits are now shared across our infrastructure (no more inconsistent counts).
• Fallback headers — responses include X-Kyma-Fallback: true when a backup was used, plus X-Kyma-Fallback-Layer: 1|2|3 indicating how deep the fallback went.
• More models available — expanded backup infrastructure means models stay available even during provider outages.

​

April 4, 2026

​

🚀 Launch

• Launch model set — Llama 3.3 70B, Llama 4 Scout, Qwen 3 32B/235B, Gemma 4, Kimi K2, GPT-OSS, Gemini
• Dashboard — API keys, usage stats, playground, model browser
• Google Sign-In — one-click login
• OpenAI compatible — drop-in replacement for any OpenAI SDK
• Tier-based rate limits: Tier 0 (free) = 30 RPM, up to Tier 4 = 300 RPM
• Streaming support
• Supabase PostgreSQL backend