Availability on EvoLink
Carbon is not yet available on EvoLink. Subscribe below to hear when you can start using it through our API.
Gemini 4 Carbon is the name of a Google model version reportedly being tested internally in Jetski. Reports focus on coding, 3D work and possible partner access. As of October 22, 2026, Google has not announced a public Carbon API, and Carbon is not yet available on EvoLink. Get notified when access opens, or start building with Gemini 4 Argon through EvoLink’s unified API.
Last checked:
| API detail | Current status |
|---|---|
| Google public API | Not listed in Google’s public model catalog. |
| EvoLink access | Not yet available. Subscribe for an availability update. |
| API model ID | Not yet published. |
| Context window | Not yet published. |
| Maximum output | Not yet published. |
| Input types | Not yet published. |
| Output types | Not yet published. |
| Reasoning controls | Not yet published. |
| Tool calling | Not yet published. |
| Structured output | Not yet published. |
| API pricing | Google and EvoLink have not published Carbon rates. |
Official availability, internal testing and community reports at a glance.
Google’s model directory currently lists Gemini 4 Argon. Carbon’s public name and release date have yet to be announced.
Google’s October 13 release notes cover Argon availability. Carbon’s API endpoint, model ID, pricing and token limits have not been published.
Carbon is not yet available on EvoLink. Subscribe below to hear when you can start using it through our API.
Business Insider’s Hugh Langley reports that Google employees are testing Carbon in Jetski. The report leaves open whether it will arrive as an Argon update or a separate model; Google did not comment.
On October 21, Lentils reported partner early access and LMArena testing. Google has not announced Carbon’s participation in its early access program, and no public Arena session or model listing is available to check the report.
Carbon’s pricing, context window, output limit and supported inputs are awaiting publication. We will add them alongside the API documentation when they become available.
We will update this page as Google publishes details and EvoLink prepares access. Integration information will include the model ID, supported features and billing.
Leave your email and we will notify you when Carbon becomes available on EvoLink.
We will only email you about Gemini 4 Carbon availability.
Follow the discussion Join EvoLink on Discord
Your update will include the EvoLink model ID, pricing and supported features so you can plan your integration. In the meantime, existing Gemini models are available through the same unified API.
Save your current results and make model selection easy to change when Carbon arrives.
Run representative coding and document tasks with Gemini 4 Argon. Keep prompts, accepted outputs, tool traces and failures for later comparison.
For code changes, check the tests and review the diff. For documents, check the citations. Record the edits a person needs to make before the result is usable.
Track response time, tokens, retries and completed tasks on your current model. When Carbon’s rates and usage are available, use the same records to compare costs.
Use EvoLink’s unified API with a configurable model choice. Try Carbon on a small set of tasks after checking its formats, tools and error handling, with the existing model ready as a fallback.
Key reports through October 21, 2026, 22:05 Beijing time. Google’s model directory and changelog were checked on October 22. All times below are UTC+8.
| Date / time (UTC+8) | Source | Update |
|---|---|---|
| 2026-10-10 03:37 | Business Insider: internal testing | BI reports employee testing of Carbon in Jetski, with its public name and release form still open. |
| 2026-10-10 04:46 | Vedant Misra: RSI response | Misra mentions recursive self-improvement in response to the coverage; training details remain unpublished. |
| 2026-10-13 | Google: Argon availability | Google expands Argon availability, making it the public Gemini 4 option for current integrations. |
| 2026-10-19 22:07 | Stephanie Palazzolo: model preview | The journalist previews a report about upcoming models, including Carbon. |
| 2026-10-19 23:34 | Journalist’s timing correction | The expected window for the model group changes from “this month” to “the coming weeks.” |
| 2026-10-21 12:59 | Lentils: partner / Arena claim | Lentils reports partner access, Arena testing, better 3D, weak UI and slower but more complete work. |
| 2026-10-21 13:00 | lyra: quoted community reaction | lyra quotes the post and uses the playful nickname “Gemini 4 Cabrón.” |
| 2026-10-21 13:31 | DC Inside: Korean repost | A Korean community post links to lyra and brings the same report to DC Inside. |
The following entries come from third-party projects. They show how those projects are preparing for possible Carbon support; Google’s API specifications are still pending.
| Field / source | Code value | Explanation |
|---|---|---|
| Model string | gemini-4-carbon | orgtree uses this provisional name and notes that the entry was copied from Barium. |
| Context value | Some(1_000_000) | The 1M value comes from the copied client configuration. Carbon’s official context limit has not been published. |
| Reasoning selector | low / high | The client tests low and high choices. Server-side reasoning options are awaiting documentation. |
| Seat / pricing object | seat: 2.0; prices: None | seat is an internal application field; prices: None leaves a separate price object empty. Token rates will come from the provider’s pricing information. |
| Conditional listing | conditional: true | The entry is shown only when the account catalog includes the model. The linked test supplies a sample list rather than an account response. |
| ORIONFIRE enum | MODEL_GOOGLE_GEMINI_ORIONFIRE = 343 | 343 identifies ORIONFIRE within this protocol. Its relationship to Carbon is still being discussed; the number carries no model-size or context information. |
| Claimed internal route | orionfire;/aistudio/gemini-v4s-jarvis-sc | The teardown author reports this internal-looking string. The original binary and Carbon association are still to be checked; public API documentation will provide the endpoint for integration. |
The story has moved from employee testing to release expectations and reported partner access.
Hugh Langley’s BI report describes Carbon in Jetski. The Decoder and Korea’s AI Times picked up the story, including employee coding impressions and Misra’s RSI response. Benchmark scores and training details were not part of those accounts.
The Information mentioned Carbon in its paid newsletter. Publicly available context comes from the journalist’s preview and timing correction; Investing.com and Korean outlets relayed that same report.
Lentils described partner access, Arena testing and mixed performance: better 3D, weaker UI and slower responses with more complete work. lyra and Korean posts spread the news. Google’s EAP exists, while Carbon’s participation is still awaiting an announcement.
The surrounding discussion includes client preparations, an Argon alias theory and English and Korean coverage.
orgtree labels its Carbon entry provisional and copies settings from Barium, including a 1M context value. The seat field belongs to the application. Official model details will determine the eventual integration.
OrcaRouter describes Arena Scout capability changes and a 3D mascot clip. The original bot record and replayable session are not linked in the available account, leaving the proposed Carbon identity open. The 3D scene may have been rendered from code or tools.
English and Korean coverage often traces back to BI, The Information or Lentils. Translations and reposts broaden the discussion but share those original reports. Video links are also circulating; the examples on this page are based on written reports and code.
This Rust excerpt shows a provisional model entry in the orgtree client.
The project notes say Barium inherited Argon’s placeholder seat, effort and usage rates, and Carbon then copied Barium. This explains the 1M context and seat values shown below; Carbon’s published specifications are still pending.
The cost estimator groups Carbon with other pro-style models and uses shared estimates. Carbon-specific official rates, billing records and a successful request are not included in this example. EvoLink will publish its pricing with verified access.
// speculative Gemini 4 Carbon (user 2026-10-10): a copy of Barium, offered
// only while the account's `agy models` lists it
Tier { tier: "carbon", provider: GOOGLE, seat: 2.0, model: "gemini-4-carbon",
versions: &[("4", "gemini-4-carbon")], letter: "C",
context: Some(1_000_000), live_effort: false, legacy: false, conditional: true, prices: None },orgtree · catalog.rs · L76 · Provisional Rust model entry.
The ORIONFIRE name in Antigravity protocol files has prompted discussion about a possible Carbon connection.
The cited file defines MODEL_GOOGLE_GEMINI_ORIONFIRE as 343. Other public copies use the same name and may share a source. The code leaves its relationship to Carbon unspecified.
A Reddit post reports orionfire;/aistudio/gemini-v4s-jarvis-sc in an Antigravity v2.25.0 binary. The original file, platform, hash and extraction offsets, and the model behind this name, are details still to be supplied.
Google’s changelog lists App v2.25.0 on October 20 and tracks App and IDE versions separately. The entry covers the application release without mentioning Carbon or ORIONFIRE. Carbon API details are still pending.
MODEL_GOOGLE_GEMINI_ORIONFIRE = 343;Antigravity-Tools-LS · exa.codeium_common_pb.proto · L276 · Protocol enum definition.
Early reports describe stronger 3D work, mixed frontend results and slower but more complete answers. Here is what each report says and what to compare in your own tasks.
An employee quoted in the BI reporting described coding with Carbon as feeling similar to Opus 5.5. That is an early user impression; a published head-to-head benchmark is still to come.
Lentils reports better 3D results than Argon while frontend and UI work remain weak. The reported gains are therefore more specific than an across-the-board improvement in visual work.
Lentils also describes Carbon as slower but less likely to skip parts of a task. For an integration, both response time and the amount of finished work will matter.
Gemini 4 Argon already has API documentation and an EvoLink integration. Use it to build and test your application while you follow Carbon’s progress.
Check the current Gemini 4 Argon page for supported API behavior and live pricing before choosing your workload.
For repeated everyday requests, evaluate Gemini 3.8 Flash on the same quality, response-time and budget requirements.
When Carbon is available, compare it with your existing model on quality, reliability and cost per completed task. Switch where the results justify it and keep a fallback for the rest.
Compare the features and pricing of the Gemini models currently available on EvoLink.



Check supported features and current pricing for Gemini 4 Argon.
Choose an existing model for your quality, latency and budget needs.
Set up authentication and your client with EvoLink’s guide for the existing Gemini API.
Carbon is a name used in reports about an internal Google Gemini 4 model version. Business Insider describes employee testing in Jetski. Google has yet to announce its public name or whether it will be an Argon update or a separate release.
Not yet. You can subscribe to an email update when EvoLink makes Carbon available. For an integration today, start with Gemini 4 Argon or one of the existing Flash models.
Lentils reported partner early access and Arena testing on October 21. Google’s general EAP offers selected developers, researchers and partners access to future models, but its public page does not name Carbon or offer a Carbon application. A public Arena session and model mapping are still missing from the report.
Google has not announced a date. On October 19, Stephanie Palazzolo changed the expected timing for a group of upcoming models from “this month” to “the coming weeks.” That is the latest wording of the report, rather than a release date for Carbon or EvoLink.
Google API access means calling the model directly through Google. EvoLink access means using it through EvoLink’s unified API. As of October 22, 2026, Google’s public catalog does not list Carbon, and Carbon is not available through EvoLink. You can subscribe here for an EvoLink availability update.
Google has not published a Carbon API model ID. The gemini-4-carbon name identifies this page; the model ID to put in an API request will be listed in the integration documentation when Carbon becomes available.
Google and EvoLink have not published Carbon API prices. Carbon’s rates will be added here when available.
Google has not published Carbon’s context window or maximum output limits. These limits will be added to the integration documentation when available.
Google has not published Carbon’s support for tool calling, structured output or reasoning controls. The supported features and request parameters will be listed in the API documentation when available.
Carbon is not yet available through EvoLink, and its supported API formats have not been announced. Google offers an OpenAI compatibility layer for supported Gemini models; Carbon’s integration details are still pending.
Google has not announced Carbon’s API access requirements. Gemini app subscriptions and Gemini API billing are separate services; a Pro or Ultra subscription does not currently provide public Carbon API access.
Argon is a publicly documented model whose availability expanded on October 13. Carbon is an internal version described in reports, with its eventual name and product tier still open. Pro, Flash and 4.1 are not confirmed names for it. The community nickname “Gemini 4 Cabrón” refers to the same discussion.
The available accounts are an employee’s Opus 5.5-like coding impression and selected community comparisons with Argon. A controlled benchmark would help establish the difference. For your application, compare the same tasks, tools and budgets, including reliability and cost per completed task.
Community posts discuss 3D scenes, which may be rendered from generated code or tools. Google has not published native video-generation features or input/output formats for Carbon.
We will email you when Carbon is ready to use through EvoLink, with integration details. This is an EvoLink notification subscription; Google partner testing has its own access process. It includes no reservation, fixed launch date or pricing preference.
Vedant Misra mentioned recursive self-improvement while sharing the coverage. Carbon’s training method has not been described publicly, so the extent of any RSI use is still open—including whether it involves weight changes or a particular training loop.
The orgtree entry is a provisional client configuration copied from Barium. It contains the 1M value and appears only if an account catalog lists the model. Official Carbon documentation is still needed for the API ID, actual context limit and token prices.
ORIONFIRE appears as an enum in third-party Antigravity protocol files. A community teardown author associates it with Carbon, although that link is not recorded in the cited code. The enum number is 343; the reported /aistudio/ path belongs to the teardown discussion, with public API details still awaiting announcement.