Skip to main content

​
Key Takeaways

  • Available now: Google announced Gemini 4 Carbon on October 28 and opened it to all paying developers and enterprises on November 10. APIYI connected on November 11. Model name gemini-4-carbon, available in the default and svip groups, on both Gemini-native and OpenAI-compatible endpoints
  • Limited-time $2 / $10 per 1M tokens through November 30, 2026. From December 1 the standard $4 / $20 applies and APIYI will follow
  • Built for the hardest tasks: ranked #1 on the Artificial Analysis Intelligence Index at 59, state of the art on three cybersecurity benchmarks, and #1 on CursorBench 4.0, WANDR and AutomationBench
  • Everything we tested passed: both endpoints, streaming, function calling with round-trips, JSON Schema and image input, with zero errors at low concurrency
  • Re-tested after the resource adjustment: every request hit gemini-4-carbon; a 199K-token document passed the needle test, and maxOutputTokens is now enforced

​
Background

Carbon is the second model in the Gemini 4 family, arriving just two weeks after Argon. Google is not pitching it as a better chat model but as a model that pushes long-horizon work to the limit: multi-package migrations, cross-subsystem root-cause investigations, complex architectural redesigns, open-ended scientific discovery and cybersecurity defense. The rollout came in two steps. On October 28 it went to a set of cyber defenders through the Fairwind Program. On November 10 it opened to all paying developers, enterprises and Google AI Ultra subscribers, and landed in Google Antigravity, the Gemini app and AI Mode in Search. APIYI connected the next day. The relationship with Argon is simple: both share the 1M window, built-in tools and per-token price; the difference is how long they are willing to think. Argon is the efficiency tier for everyday code and knowledge work. Carbon is the hard-problem tier: on the same task it spends more thinking tokens exploring, verifying and checking edge cases, so it is slower and more expensive, with a higher ceiling.

​
Detailed Analysis

​
Official benchmarks

All figures below are published by Google and were not independently reproduced by APIYI: Third-party Artificial Analysis separately reports two numbers worth knowing: an AA-Omniscience hallucination rate of about 11% (Argon: about 4%) and a very verbose output profile (about 200M output tokens for a full index run versus a 64M median). The ceiling is higher, but factual work still needs sources and verification.

​
Cybersecurity is the main thread

Google splits Carbon’s cybersecurity capability into find, validate and patch, with a benchmark for each:
  • Find: SEC-bench Pro 88.2%, autonomously locating complex flaws in real software.
  • Validate: ExploitGym 54.8%, deciding which findings are actually exploitable so teams can prioritize.
  • Patch: CWE-bench 71.0%, turning validated findings into working fixes.
Google’s internal examples claim 100% scan recall and 95% triage precision in Project Fortify, and a more than 25% relative recall improvement over Argon on open-source code tests. These are Google’s figures, and the cybersecurity benchmarks were run without guardrails; real deployments will be shaped by safety policy.

​
Long-horizon work and multimodality

Carbon’s other pitch is endurance. Google says internal teams ran it for days on multi-step jobs, root-causing and patching subtle bugs across large codebases. On WebDev Arena it gained 89 Elo to reach #4, and for frontend work it runs build-and-verify loops, recovers from errors and checks both the code and the rendered UI. The third thread is multimodal knowledge work: synthesizing grounded answers across dense PDFs, spreadsheets and slides, and following schemas strictly to produce spreadsheets, presentations and reports. On the new MKA benchmark co-developed with Artificial Analysis, Carbon ranks #2 with more than 6x lower cost per task than the top model.

​
API changes

  • 1M unified window: input, thinking and output share 1,048,576 tokens; up to 65,536 tokens per streaming segment, with long-decode continuation beyond that.
  • Thinking levels: medium by default, with low / medium / high. thinkingBudget is no longer supported and should return 400; minimal is also unsupported.
  • Fixed sampling: temperature 1.0, topP 0.95, topK 64; custom values are ignored.
  • Caching: implicit caching only (4,096-token minimum); explicit cachedContents is not supported.
  • File input: besides PDF and plain text, Carbon accepts docx, pptx, xlsx and other Office / OpenDocument files.
  • Knowledge cutoff: March 2026.

​
Pre-launch test

We ran a low-concurrency test on APIYI covering both endpoints, streaming, function calling, JSON Schema, images, thinking levels and a long-document needle test. Full data is in the test report; the conclusions are here. What passed: non-streaming and streaming work on both endpoints; function calling correctly emits get_weather({"city":"Beijing"}) and consumes the tool result; JSON Schema returns valid JSON on both; image input recognized a Chatbox settings screenshot; thinkingLevel moved thinking from 126 to 1,404 tokens between low and high. Low-concurrency latency: 60 short requests with zero errors, P50 12.6s; the long tail still reaches 60–170 seconds, and one call timed out at 180 seconds. Set client timeouts generously (Google recommends 15–20 minutes for complex work) instead of cutting off early. What changed after the re-test (details in the usage guide):
  1. Provider-side model switching: some responses previously reported 3.1-flash-lite-05-2026; it did not reproduce after the resource adjustment. Google retains the ability to switch the answering model per request, so log modelVersion and retry if it happens; it does not mean a request violated policy.
  2. 1M context: a 199K-token document passed the needle test at promptTokenCount=199,277; the earlier ~128K cap and silent truncation did not reproduce. We did not test 500K / 1M this round; validate with a small sample first.
  3. maxOutputTokens is enforced now: it includes thinking tokens. Sending 50 / 100 / 300 all exhausted the budget during thinking, returned MAX_TOKENS and no answer; set a few thousand for hard tasks.

​
Pricing and availability

  • Model name gemini-4-carbon, available in the default and svip groups.
  • The limited-time rate runs through November 30, 2026; from December 1 it follows Google’s standard price.
  • Native endpoint: POST /v1beta/models/gemini-4-carbon:generateContent; OpenAI-compatible endpoint: POST /v1/chat/completions.
Model pricing is aligned with the official site and may change; the table above is for reference only. The Model Pricing tab in the top navigation is authoritative: Model Pricing.

​
Summary and recommendations

Carbon is for hard problems: long-horizon coding, cross-system root-cause work, scientific reasoning, security auditing and dense multimodal synthesis. For everyday code and interactive tasks, Argon is faster and cheaper; Carbon’s extra thinking may not pay off. At launch, we recommend:
  • Log modelVersion for critical work and retry if it reports Flash-Lite; add verification for consistency-sensitive flows.
  • Validate very long documents with a small sample first (199K tokens verified); chunk when going larger.
  • Leave enough maxOutputTokens headroom so thinking does not consume the whole answer.
  • Use low for simple tasks and high for hard ones; thinking tokens are billed as output, so high costs noticeably more.
The model is callable on APIYI today with the same integration pattern as other Gemini models. Full parameters, limits and examples are in the Gemini 4 Carbon guide.