Get a key

Corent

Changelog

What shipped, when. Everything here is live: the roadmap earns its place on this page only by shipping.

August 7, 2026

A 30-second single take

  • Seedance 2.5 is live: 30 seconds of video in one generation at 720p, double what 2.0 could do and the longest single take in the catalog. It trades resolution for length, so shorter clips still reach higher resolutions on other models.
  • It is pinned by name (corent-seedance-2.5) rather than reached through a tier. A tier is a promise about quality, and this model is too new to carry an independent quality score yet.
  • Its price is measured, not estimated: this model bills by token rather than by second, so the per-second rate was derived from real renders and rounded up, which can only cost us and never you.

August 2, 2026

Twice the throughput, and money that cannot be spent twice

  • The API now runs two workers on twice the memory, roughly doubling how many requests it serves at once.
  • Text requests place a worst-case fund hold before any provider is called, and every exit path returns it (settled, failed, disconnected, or timed out). Concurrent calls can no longer read the same balance and all pass.
  • Every published price now clears a floor above what the job actually costs to run. 82 of 141 menu prices moved as a result, each by exactly one cent.

July 29, 2026

Held money, in plain sight

  • Your balance now separates held from available. While a generation is in flight, the dashboard shows how much is reserved against it and how much you can actually spend, and the generations-remaining estimate reads available rather than total.
  • Money is reserved before a provider is paid and released on every exit path, so an in-flight job can never quietly hold your balance hostage after it finishes.
  • One published name per model: corent-*. Receipts, job polls and the speech endpoint all answer in that single spelling. Every older name you may already be sending still works.
  • The routing brain's measured quality scores now actually reach production. 34 models get a measured number for the first time, 16 of them text models, which matters now that text routes through the same brain.

July 27, 2026

Sign in with Google, and shapes that stop lying

  • Log in with Google, on both the login and signup pages. A failed or cancelled sign-in now explains itself on our own page instead of dropping you on a blank form.
  • Aspect ratio no longer silently defaults to 16:9 on video. Name a shape a model cannot render and the request is refused before any charge; omit it on image-to-video and the clip follows your source image instead of being reframed.
  • Clip length now resolves to the model's own default, snapped onto the durations it really offers. What we quote, what we bill and what we render are one number.
  • GET /v1/models publishes only the shapes you can actually ask a model for.
  • The site stopped quoting per-model prices. Tier prices on /pricing keep rendering live from the API.
  • Text generation now routes through the same scoring brain as image and video.

July 25–26, 2026

Direct model access, and an OpenAI-compatible text endpoint

  • New model parameter on image, video and speech: pin an exact catalog model and it is never routed or substituted. If that model cannot do the job, you get an honest failure instead of a quiet swap.
  • GET /v1/models: the public menu of every live model with quality scores and capabilities. No auth needed. Model names are public; the suppliers behind them are not.
  • Per-key model access: an API key can be full access, tiers only, or locked to a named set of models.
  • POST /v1/chat/completions: OpenAI-compatible text generation, Corent's fourth category. The model field takes a tier (corent/text-premium, routed) or an exact model (pinned). SSE streaming carries token usage and the exact cost in the final chunk, and tool calling and JSON mode pass straight through.
  • SDKs 0.2.0 for Python and TypeScript, plus the MCP server: the model parameter and a models() catalog call.
  • Receipts on a pinned job name the model you pinned, never the route that served it.

July 22–23, 2026

New models at the top of the board

  • Gemini Omni Flash is live for video, confirmed by a proof render measured at 1280x720, 24fps, 10 seconds. Nothing goes live here on a press release.
  • GPT Image 2 now runs through OpenAI's own door, and Veo 3.1 through Google's, with a lite variant of Veo added alongside.
  • MAI-Image-2.5 and Nano Banana 2 Lite added to the image tiers.
  • A second voice supplier, Gemini TTS, joins the speech lane.
  • Sora 2 removed: OpenAI deprecated it, its API shuts down in September 2026, and its standing had already dropped below the tier bar.
  • Voice price corrected from 27¢ to 14¢ per 1,000 characters. The old figure rested on a legacy credit estimate and had us charging 2.7x the direct rate for the same model.
  • Dashboard usage history now pages through every generation, 20 per page, with the prompt and a link to the output on each job.

July 21, 2026

Official SDKs, honest price ranges, measured receipts

  • Official SDKs: pip install corent and npm install corent-sdk: three lines to a render, with automatic retries, idempotency keys, and job polling built in.
  • Images accept async: true and return a job id immediately, so a client timeout can never lose a finished render. The SDKs use this by default.
  • Receipts now state the MEASURED pixels of the delivered file for video as well as images, parsed from the file itself, never echoed from the request.
  • The tier catalog and pricing pages show honest min–max price ranges across each tier's live models instead of a single average; any charge inside the range is normal routing.
  • Routing brain: the vision inspector's quality scores now feed model selection per workload, wrong-shaped deliveries penalize the responsible model automatically, and a misbehaving model is benched individually instead of its whole provider.

July 20, 2026

Voice goes public + billing hardening

  • Voice is now a first-class product: POST /v1/audio/speech (text to speech, billed per 1,000 characters) joins image and video in the docs, the MCP server, and the public tier catalog.
  • Video ladder rebalanced so both price axes rank the same way: premium now estimates from $0.59 per clip (was $0.71), pro from $0.87 (was $0.80). Per-second rates now rise tier by tier.
  • Batch items accept tier and resolution, matching the single endpoints.
  • Crypto deposits documented in the public API spec (POST /v1/account/deposit/crypto).
  • Billing hardening: output is validated before any charge on every path (an empty file can never bill), settlement can never exceed the duration you consented to, and concurrent retries with the same Idempotency-Key can never charge twice.
  • Docs prices now render from the live tier catalog, so the documentation can no longer drift from what the API actually charges.

July 17–18, 2026

Five tiers, one knob for resolution

  • New tier system across image and video: air, lite, premium, pro, max_pro. One vocabulary from the API to the pricing page. The legacy preference field (fast/cheap/balanced/quality) still works and maps into the same tiers.
  • tier parameter on the generate endpoints (preferred over preference).
  • resolution parameter on video: 720p, 1080p, or 4k. Pro allows up to 1080p, max_pro up to 4K; asking above a tier's cap clamps down instead of erroring.
  • GET /v1/tiers now lists per-resolution prices per tier, including per-second video rates.
  • The pricing page renders from the live tier catalog.
  • MCP server 0.4.x: five-tier and resolution parameters on the generate tools.

July 14, 2026

Receipts API, tier catalog, idempotency, provenance plumbing

  • GET /v1/requests: your receipts, with the per-request exact charge, tier, candidates considered, failover, billed true/false.
  • GET /v1/tiers: public machine-readable catalog of every live tier with price estimates and live status. No auth needed.
  • Idempotency-Key header on generate endpoints: a retried request returns the original job, so double-charging by retry is now impossible.
  • New API keys start with a $10/day cost cap by default (raise it anytime via the dashboard or the limits API).
  • Batches now count as one request per item against all rate and spend limits.
  • Content moderation now guards every generation path, not just the demo.
  • The router now logs real outcome signals per generation, so routing quality starts learning from live traffic.
  • C2PA Content Credentials signing built into the media pipeline (rolling out with production signing certificates).
  • MCP server v0.3.0: structured tool results, read-only tool annotations (fewer confirmation prompts in agent clients), machine-readable error codes.

July 13, 2026

Billing precision

  • Fixed a rounding edge case that could overcharge one cent on some tier prices.
  • Billing now settles only after your file is safely stored: a failed delivery can never bill.
  • In-flight video now counts toward daily spend caps while it renders.

July 12, 2026

Image tier reshuffle at unchanged prices

  • Fast tier pool upgraded at the same 2¢ price; the router picks per prompt, speed first.
  • Premium tier led by the current arena-leading model at 6¢.
  • Video pricing and tiers unchanged.

July 11, 2026

Smarter routing + email system

  • Per-skill quality ratings: the router now scores every model per prompt type (photos vs logos vs text-in-image).
  • Quality floor: cheap and fast preferences only choose among models that clear a quality bar for your specific prompt.
  • Transactional email live: deposit receipts and low-balance warnings.

July 9, 2026

Agent-native release

  • MCP server published on npm (corent-mcp) and hosted at mcp.corent.tech/mcp.
  • Natural-language intent API: describe what you want, Corent plans it, you approve the cost, it generates.
  • Batches (up to 50 items), signed webhooks with delivery logs, and SSE job streaming.
  • Crypto deposits (USDT, USDC, BTC and more).