GPT-6.1 Sol vs Gemini 4 Carbon: Five Index Points, Six Times the Bill, and One You Can't Call

The comparison between GPT-6.1 Sol and Gemini 4 Carbon has an unusual shape: one of them has been answering production requests since 29 September 2026, and the other has been answering benchmarks since 28 October. Google's Carbon is real, it is the highest-scoring model on the independent board, and you cannot send it a request unless you are a cybersecurity defender admitted to Google's Fairwind program. No public model string, no published price page, no endpoint. GPT-6.1 Sol you can call this afternoon. That asymmetry does not make the comparison pointless — it makes it a different comparison from the one the leaderboards imply, and a much more practical one.

One is a product, the other is a programme

Google announced Gemini 4 Carbon as "our next frontier model" on 28 October 2026, two weeks after Gemini 4 Argon. Access opened to a subset of cyber defenders through Fairwind, with broader developer, business and consumer access promised "as soon as possible" and no date attached. As of 5 November there is still no generally available model identifier. The pricing Google has stated — $2.00 per million input tokens and $10.00 per million output, cached input 95% off, rising to $4.00 and $20.00 at standard rates with no confirmed end date for the introductory tier — is a rate card for a service that most readers cannot buy.

GPT-6.1 Sol has been generally available on the OpenAI API since 29 September at exactly the same headline rate: $2.00 in, $10.00 out, $0.10 cached input, $2.50 cache write, with a long-context step that reprices the whole request at $4.00 input and $15.00 output above 272,000 input tokens. Our catalogue carries it as openai/gpt-6.1-sol, listed 2026-09-29, at OpenAI's own price with no markup. We do not carry Gemini 4 Carbon, because there is nothing to carry yet; we carry Gemini 4 Argon.

So the honest framing is not "which should you switch to". It is "what does Carbon's five-point lead actually consist of, and what would you pay for it if it were on the menu".

Identical rate cards, and a bill that diverges by six times

Both models list at $2.00 and $10.00. Read that line and you would conclude they cost the same to run. The independent measurement says otherwise, and the reason is volume rather than price.

• Intelligence Index v5.0.1 — Gemini 4 Carbon 60 vs GPT-6.1 Sol 55, both at maximum published effort

• Output speed — Carbon 68 tokens/sec vs Sol 56 tokens/sec

• Cost per completed index task — Carbon $7.76 vs Sol $1.27

• Output tokens burned on the index run — Carbon 205 million vs Sol 63.7 million

• Time per index task — Sol 20.3 minutes, with 157.57 seconds to its first answer token

• Context window — Carbon 1M vs Sol 1,050,000

• Maximum output in one response — Sol 128,000 tokens; Google has not published a Carbon figure

• Modalities — both accept text and images and return text

• Availability — Carbon gated to Fairwind defenders, no public identifier vs Sol generally available since 29 September

• List price — both $2.00 input and $10.00 output per million, before Carbon's move to $4.00 and $20.00

Carbon produced 3.2 times Sol's output tokens and cost 6.1 times as much per task. The bill grows faster than the token count, which is what happens when a model thinks longer and ingests more context per attempt. A five-point index lead bought with six times the spend is a defensible trade for a security team running a few hundred high-stakes analyses; it is an indefensible one for anything with a queue.

Where Carbon is genuinely ahead

It would be easy and wrong to file Carbon as an expensive tie. On the evaluations Google built or selected for, it leads clearly, and those are the evaluations closest to the announced use case.

• SEC-bench Pro, vulnerability detection — Carbon 88.2%. Google-reported.

• ExploitGym, exploitation — Carbon 54.8%. Google-reported.

• CWE-bench, patching — Carbon 71.0%. Google-reported.

• Gray Swan indirect prompt injection — 0.0% attack success rate. Google-reported.

• GDPx, SemiAnalysis's 206-task professional benchmark — Carbon about 10.9% at roughly $16 per task, behind Claude Opus 5.5 at about 12% and roughly $51 per task.

• AA-Omniscience — Carbon leads the knowledge-accuracy board; independent coverage has put its hallucination rate near 11%, up from Argon's figure, which is worth knowing before treating a cyber lead as a general reliability lead.

The cyber numbers are the reason the model exists and the reason it is gated. They are also unreproduced by anyone outside Google, run without guardrails according to the launch coverage, and published without a separate system card at launch. Label them as vendor-reported and move on; they are not the reason a general engineering team would pick this model.

The lead is domain-shaped, not general

The most useful third-party result of the past week is a negative one. When SemiAnalysis re-ran its GDPx ranking with cybersecurity, biology and alignment tasks excluded, the order changed: GPT-6 Astra moved to first, and both GPT-6.1 Sol and Claude Sonnet 5.5 moved ahead of Gemini 4 Carbon. SemiAnalysis said the revised ordering better matched its own subjective impressions.

That is a precise statement about what Carbon is. It is not uniformly smarter; it is disproportionately strong on a domain cluster that happens to be where Google aimed it, and the aggregate index rewards that cluster. If your workload is security analysis, that is the whole argument for waiting for access. If your workload is a document pipeline, a coding agent or a support flow, the five-point gap is largely composed of tasks you will never run, and you are comparing a model you can call against one you cannot.

What Sol does better, on OpenAI's own numbers

OpenAI's launch claims for GPT-6.1 Sol are vendor-reported and should be read as such, but they are specific enough to check later. On DeepSWE v1.1 it reports 75.22% at high effort and 71.90% at max — the unusual detail being that this benchmark peaks at high, not max. On OSWorld 2.0, 71.42% at max. On Terminal-Bench Science 0.1, 57.02% at max, against GPT-6 Astra's 68.1%. On AutomationBench 1.0.6, 36.10% at max. On GDP.pdf, 32.0% at high, level with Astra.

The pattern in those numbers is the useful part: Sol is close to Astra on document and coding work and clearly behind it on science and long-horizon automation. It is a workhorse with a frontier-adjacent ceiling, sold at a fifth of Astra's $10.00 and $50.00. Independent measurement agrees on the shape — 55 on v5.0.1 is not a frontier score, but $1.27 per completed task is the lowest in this comparison by a wide margin.

The decision, for the two readers this article has

If you are a defender inside Fairwind, the trade is real and the numbers above are the trade: five index points and a large cyber-specific lead, for six times the cost per task and a rate card that doubles when the introductory period ends. Run Carbon where the analysis is the deliverable and Sol where the volume is.

If you are everyone else, Carbon is not a decision, it is a watch item. The model to compare against today is Gemini 4 Argon, which our catalogue carries as google/gemini-4-argon at the same $2.00 and $10.00 introductory rate, and which scores 50 on the same v5.0.1 index — five points below Sol, at 68 tokens per second and a cost per task in the same range as Carbon's. Argon doubles to $4.00 and $20.00 at standard rates, which is the date to put in a calendar.

The reason to care about the gating rather than just the scores is that a routing layer makes the eventual answer cheap. Both Sol and Argon sit behind one key here, at provider list price with no markup, so the comparison is a config change rather than a second contract. When Carbon does open, the routing DSL is where a per-task split belongs: cyber analysis to the specialist, everything else to the workhorse, with failover if the specialist is rate-limited on day one. Building the pipeline against one model and hard-coding it is the decision that ages badly, and this pair is the clearest argument for not making it.

Five points is a real difference between two models. It is not, on its own, a reason to rebuild a stack around a service that has not opened its doors.