
Gemini 4 Carbon Went to the Lab. Google's Own Write-Up Credits Gemini 4 Argon Thirteen Times
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 85 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 127 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 45 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 194 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 136 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 123 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 182 tok/s
Fifty million lines of Lean 4. That is what it took to machine-check a proof of the Courtade–Kumar conjecture, a 2013 result in information theory about how much of a noisy bit string survives compression, and it is the largest single artefact in the research write-up published on 13 November 2026 under the title "Advancing science & engineering with Antigravity's /teamwork." The post is the source of the claim now circulating that researchers are putting Gemini 4 Carbon to work on multi-step problems across physics, mathematics and the life sciences. Carbon is in the post. It is named exactly once, in a list of three models at the top. Every one of the case studies underneath — the conjecture, the formalised Weil conjectures, the NMR structure elucidation, the Go engine, the Zircon kernel rewrite — is attributed in the body to Gemini 4 Argon, or to "Gemini 4" without a tier, or to the harness itself.
That is not a scandal and it is not a correction. It is a distinction worth getting right before anyone budgets against it, because the two models are not interchangeable in the way the tweet implies, and because Carbon's price is scheduled to double on 1 December 2026 — fourteen days from today. What follows is what the post actually says, which claims survive scrutiny, which are vendor-reported and unreproduced, and what a reader outside Google can do with any of it this week.
The post, the programme, and the one sentence
Google Research published it on 13 November, credited to Vahab Mirrokni, a Google Fellow and VP of Google Research. The opening paragraph names the models in a single list: "Using Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen and multi-agent workflows such as Antigravity's /teamwork (building on various Gemini models), these collaborations have produced new innovations…" That sentence is the whole of Carbon's presence in the argument. Count the names in the body below it and Argon appears thirteen times, Hydrogen once, and Carbon not again.
The reason is structural rather than embarrassing. /teamwork is a harness, not a model: it runs a pool of specialised agents, gives each one only the files and tools its step needs, hands work to fresh successors so context does not rot over multi-day runs, and routes each prompt to a workflow pattern chosen for the task. Its verification layer is what Google is actually selling — the Lean 4 checker, unmodified test suites, execution simulators, hardware counters — acting as falsifiers that decide which candidate answer survives. When a harness does the orchestrating and a proof checker does the judging, the model underneath matters less than the loop around it. Argon, the model that has been callable on the paid API since 13 October 2026, is the one the campaigns ran on.
Carbon's own case for being in the list is real but different: it is the strongest general-purpose model Google has published, and on Artificial Analysis it currently sits fourth of 81 models at 60 on Intelligence Index v5.0.1. Nothing in the 13 November post contradicts that. Nothing in it demonstrates Carbon doing the work either.

The results, ordered by how checkable they are
The post is unusually generous with specifics for a research blog, which makes it possible to sort the claims by what kind of evidence backs them rather than by how impressive they sound.
• Machine-checked, so verifiable in principle — the Courtade–Kumar proof, formalised end to end in Lean 4 at 50 million lines, with the human contribution named: a CUHK team proposed the candidate function that reduced the problem to a 14-variable optimisation. Also in this bucket, the formalisation of the Weil conjectures at roughly 1.8 million lines across 28,800 subagent invocations, Hilbert's tenth problem over general rings at 919,000-plus lines, the Birkhoff conjecture open since 1927, and the 3D Bochner–Riesz conjecture, where the system reports finding and fixing a gap in a recently released informal proof. A Lean proof that compiles is a different class of claim from a benchmark score, and Google deserves the credit for that — though "50 million lines of generated proof code" is also a statement about search volume rather than elegance.
• Measured, with the measurement described — the molecular chemistry result. Thirty specialised agents in a self-correcting loop, plus a physics-informed spectral simulator, reached 95% top-1 exact match and 100% top-3 on NMR structure-elucidation benchmarks, which Google describes as nearly double single-agent performance. This is the closest thing in the post to a life-sciences claim, and it is the one most likely to be quoted. The benchmark molecules are not named in the post, the baseline single-agent number is given only as a ratio, and no independent replication exists yet.
• Peer-adjacent, with a survey attached — Deep Review at STOC 2027. Authors of 578 submissions, more than a third of the total, opted in for automated feedback on their drafts. In an ongoing survey of 70 authors, 90% rated the feedback mostly or almost all factual, 90% said they would use it again, over 30% said it caught substantive proof bugs, and 14% reported technical gaps that took more than an hour to fix. Those are self-selected respondents describing their own submissions, which is weaker than a controlled study and much stronger than a testimonial.
• Engineering numbers from production systems — a Rust rewrite of about 800,000 lines for the Zircon microkernel passing its unit tests and booting on arm64; libgav1 ported to roughly 170,000 lines of Rust at 100% bit-exact parity and ending 20% to 30% faster than the original C++ and hand-written assembly; TPU kernel work reporting a 4.09x geometric-mean speedup across 17 JAXBench kernels; Connected Components scaling to trillion-edge graphs at 6x lower RAM; and a 9x9 Go engine that gained about 2,600 Elo in under eight days, landing between KataGo at 16 and 32 visits. These are internal measurements the company chose to publish, with no external audit and, in most cases, no released artefact.
• Not in the post at all — anything about cost. No campaign budget, no token count, no per-result bill appears anywhere in the write-up. For a system whose output includes 50-million-line proofs and eight-day autonomous training runs, that is the omission a reader planning their own campaign would most want filled.
Read the list again with the attribution in mind and the shape becomes clear: the harness produces the result, the verifier certifies it, the human experts frame the problem, and the model is the engine — one engine among three, and the post names the one it leaned on.
Why a harness result is not a model result, and how to price the difference
The incentive to conflate the two is obvious, because a benchmark score is a number you can compare and a harness result is a story about a workflow. Carbon's comparable number, measured by a third party with no stake in the harness, is 60 on Intelligence Index v5.0.1 — fourth of 81, with Claude Fable 5.2 at 65 above it and GPT-6.1 Sol at 55 below. The same page puts Carbon's cost at $5.48 per completed index task and its time at 33.2 minutes per task, both well above the medians for its price class, because it writes 205 million output tokens to finish what the median model finishes in 78.6 million.

That verbosity is not a defect in a proof search. When the task is exploring a parameter space and checking candidates against Lean, writing more is the method. When the task is a support ticket, it is the bill. The 13 November post is evidence for the first case and says nothing about the second — and any team reading it as "Carbon is the model for our agent" is borrowing a result from a harness they are not running, on a task type they may not have.
What is transferable is the design pattern, and it is not Google's alone: parallel exploration, an external falsifier that has no opinion, fresh context per step, and a critic that never sees the producer's reasoning. Every one of those is a property of how you call a model, not of which model you call.
What an outsider can actually get, and what it costs after 1 December
The post ends with the most actionable paragraph in it: external scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity and Deep Review today, and Google has opened applications giving 10,000 academic researchers six months of free access to its highest usage tier. That programme is the real news for a lab. It is also the only route on offer that is free, and it is capped, application-gated, and time-limited.
Everyone else buys tokens. Gemini 4 Carbon became generally available on the paid Gemini API on 10 November 2026, thirteen days after Google announced it, at $2.00 per million input tokens and $10.00 per million output tokens — a rate that holds through 30 November and becomes $4.00 and $20.00 on 1 December, with cached input moving from $0.10 to $0.20 on the same date. There is no free tier for the model; every processed token is billed, thinking tokens included. A six-week research campaign priced at today's rate is a campaign priced at half its December cost, which is either an argument for starting now or a reason to model the bill at the higher number.
Carbon is routed on OrcaRouter as google/gemini-4-carbon at Google's list price with 0% markup, on the same key and the same OpenAI-compatible endpoint as the rest of the catalogue, which matters here for a mundane reason: the 1 December change is a vendor price change, and a pass-through rate means the number on our side moves with Google's on the day rather than after a repricing cycle. The model page carries the current envelope — 1,048,576-token context, 65,536-token maximum output, text, image, audio, video and PDF in, text out, native tool calling and structured outputs — which is the spec sheet Google's own launch post did not publish.

The question the post leaves open
Google has now published, in twenty days, an announcement that put Carbon first on a neutral index, a general-availability post that made it callable and dated its own discount, and a research post that used it in a list beside two siblings. What it has not published is a single result it attributes to Carbon alone, at a stated cost, on a task a third party can re-run. Until that exists, the honest summary of 13 November is narrower than the tweet and more useful: the harness works, the verifier is the interesting part, and the model named thirteen times is the one the campaigns actually ran on.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
