A title card reading 'Same price for fourteen more days' with the subtitle 'Gemini 4 Carbon vs GPT-6 Sol', above a two-column panel: on the left, Gemini 4 Carbon at $2.00 input and $10.00 output per million tokens until 30 November, then $4.00 and $20.00; on the right, GPT-6 Sol at $2.00 and $10.00, with a long-context step to $4.00 and $15.00 above 272,000 input tokens.
Guides & Insights

Gemini 4 Carbon vs GPT-6 Sol: Fourteen Days of Parity, Then a Twofold Gap

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Today, Gemini 4 Carbon and GPT-6 Sol have the same headline price: $2.00 per million input tokens, $10.00 per million output tokens, $0.20 or less for cached input. On 1 December 2026 that stops being true. Carbon's rate is introductory and steps to $4.00 and $20.00, while Sol's $2.00 and $10.00 has no expiry printed on it. So the comparison a buyer actually needs is not "which model is better at the same price" — it is what the sixteen-point gap on the independent index is worth, and whether it is worth paying twice as much for from the first of next month.

There is a third model in this decision that neither vendor's marketing puts there. GPT-6 Sol is now marked deprecated by Artificial Analysis, superseded inside the same line by GPT-6.1 Sol, which scores 55 rather than Sol's 44 at the same $2.00 and $10.00. A reader choosing between Carbon and Sol in mid-November is choosing between a current model and a previous one, and the honest version of this page has to say so in the second paragraph rather than the last.

The two rate cards, line by line, before the calendar moves

Start with the rows where the two models genuinely match, because they are more numerous than the marketing on either side suggests.

• Standard input price — Carbon $2.00 per million; Sol $2.00 per million. Identical today, and Carbon doubles on 1 December.

• Standard output price — Carbon $10.00 per million; Sol $10.00 per million. Same, with the same expiry on Carbon's side.

• Cached input — Carbon $0.10 today and $0.20 from 1 December, a 95% discount either way; Sol $0.20, a 90% discount. Carbon's cache is the cheaper of the two on both sides of the date.

• Context window — Carbon 1,048,576 tokens; Sol 1,050,000. Round both to a million and stop worrying about it.

• Maximum output in one response — Sol 128,000 tokens; Carbon 65,536. This is the largest real spec difference and it only bites when you ask for a long artefact in a single turn, which is exactly what long-horizon agent work tends to do.

• Input modalities — Carbon text, image, audio, video and PDF; Sol text and image, with documents accepted on our catalogue row though the GPT-6 Sol model page lists text and image. Carbon is the only one of the two that can be handed a recording or a clip.

• Reasoning control — Sol's effort ladder runs none, low, medium, high, xhigh, max, so it can be told not to think at all; Carbon is published as low, medium and high rows, with no off position listed. On a cheap classification path that difference is the whole bill.

• Knowledge cutoff — Sol 20 April 2026; Google has published no cutoff for Carbon. One-sided, and worth knowing before building a recency-sensitive pipeline on an assumption.

• Availability — both callable today on the paid API; Carbon has no free tier and bills every processed token, thinking included, while Sol is priced across Standard, Batch, Flex, Fast and Ultrafast service levels.

The price rows are the ones that decide this comparison, and they are the ones that change without either model changing at all.

Where the sixteen points come from

Artificial Analysis measures both models on Intelligence Index v5.0.1, which folds nine evaluations — AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt v2 and Terminal-Bench-Science — into one number. That shared scoring is what makes the comparison legitimate rather than a press-release mash-up.

• Intelligence Index — Carbon 60, ranked 4th of 81 models; GPT-6 Sol 44, ranked 34th of 81. For reference, GPT-6.1 Sol sits at 55, twelfth.

• Cost per completed index task — Carbon $5.48; GPT-6 Sol $2.47. The cheaper model per token is the cheaper model per job, by 2.2×.

• Time per task — Carbon 33.2 minutes; Sol 7.3 minutes, against a median of 13.5 for the class.

• Output speed and latency — Carbon 62 tokens per second with 72.88 seconds to the first answer token; Sol 123 tokens per second with 36.21 seconds to first token.

• Token volume — Carbon emitted 205 million output tokens during the index run from 7.62 billion input, described by the tracker as very verbose; Sol emitted 65.1 million from 4.75 billion, described as fairly concise. The median for the class is 78.6 million output tokens.

That last row explains the cost row, and it is the single most useful thing on either page. Carbon is not charging more per token to do the same work; it is writing three times as many tokens to reach the answer, on a benchmark where the answer is a document or a patch rather than a proof search. A model that thinks longer in prose is being paid by the word, and the index run is where you can see the exchange rate.

One caveat belongs on the Carbon number: the tracker publishes Carbon as three separate rows, low, medium and high effort, and only the high row reaches 60. A single "Carbon scores 60" line is true at maximum effort and false at the setting most people will default to.

Screenshot of the Artificial Analysis model page for Gemini 4 Carbon, showing an Intelligence Index of 60 with a rank of 4 out of 81, 62 tokens per second output speed with 72.88 seconds to first answer token, 33.2 minutes per Intelligence Index task, $5.48 cost per task, $2.00 input and $10.00 output per million tokens with a 95% cache discount, 205 million output tokens and a 1M-token context window.

The 272,000-token line, which decides the bill more than the price does

Neither model charges extra per token past a boundary in the naive sense; both reprice the entire request. The difference is that only one of them has a boundary.

GPT-6 Sol bills $2.00 input and $10.00 output up to 272,000 input tokens. Above that, every token in the request — not the excess, the whole thing — is billed at $4.00 input and $15.00 output. Google publishes no length tier for Carbon: the same per-token rate applies whether the request is 2,000 tokens or 900,000.

Put a 400,000-token document analysis through both, at Carbon's post-December rate:

• Gemini 4 Carbon — 400,000 tokens at $4.00 per million is $1.60 for the prompt. Under today's introductory rate the same request is $0.80.

• GPT-6 Sol — the same 400,000 tokens cross the line and reprice to $4.00 per million, also $1.60. Had the request stayed under 272,000 tokens, it would have cost $0.80.

Add 20,000 output tokens and the gap opens: Carbon bills them at $20.00 per million, or $0.40; Sol at $15.00, or $0.30. So the fixed-price advantage Carbon advertises evaporates above the threshold on input and inverts slightly on output. The claim that Carbon is cheaper for long documents is true against GPT-6 Astra, which doubles to $20.00 and $75.00, and false against Sol.

Where Carbon's no-tier structure genuinely wins is predictability rather than price. A pipeline with variable document length has one rate to model and no cliff to trip over, and the cliff is easy to trip: a 300,000-token request does not cost a little more than a 270,000-token one under Sol, it costs double on input for all 300,000 tokens. If your retrieval step occasionally returns too much, that is a budget incident rather than a slow drift.

Which model is better at what, stated as narrowly as the evidence allows

Neither model has been run head to head on a task suite chosen by someone with no stake in either. What exists is the shared index and vendor material, and the index's nine components can be read individually.

• Long-horizon reasoning and scientific work — Carbon's advantage. It is the model Google put beside Argon and Hydrogen in its 13 November research write-up, and its index lead is concentrated in the frontier-reasoning evaluations rather than spread evenly.

• Agentic coding and terminal work — closer than the overall score suggests. Carbon's lead is general-purpose rather than specifically agentic, and Google's own launch material for the Gemini 4 generation has consistently shown the coding-agent numbers trailing the general index.

• High-volume production paths — Sol, or better, its successor. Half the latency to first token, four times the throughput per task-minute, and a documented off switch for reasoning. For work where the answer is short and the volume is high, Carbon's verbosity is a tax with no matching benefit.

• Multimodal input — Carbon, without contest. Sol takes text and images; Carbon takes audio, video and PDFs as well, which is the one row where the two are not in the same product category.

• Long single artefacts — Sol, on the ceiling alone. 128,000 output tokens against Carbon's 65,536 means anything that must be produced in one pass has a hard limit on the Google side.

• Continuity — Sol, but only through its successor. A team that builds on GPT-6 Sol today is building on a model the tracker already flags as deprecated in favour of GPT-6.1 Sol, which costs the same and scores eleven points higher. That is not a reason to avoid the vendor; it is a reason to specify gpt-6.1-sol rather than gpt-6-sol when the request body is written.

Running the comparison instead of reading it

Both models are routed on OrcaRouter at their vendors' own list prices with 0% markup, as google/gemini-4-carbon and openai/gpt-6-sol, on one OpenAI-compatible endpoint and one key — alongside openai/gpt-6.1-sol, which is the row the deprecation note makes relevant. That makes this particular comparison unusually cheap to settle empirically: the same request body can be pointed at Carbon, at Sol and at Sol's successor, and the only thing that changes is the model string.

Two practical notes for anyone doing that. First, the Carbon price change on 1 December is a vendor change passed through here on the day, so a cost comparison run in November and repeated in December will differ for reasons that have nothing to do with the model. Second, if the workload is exploratory — trying Carbon on real traffic before committing a production path to a model whose per-task cost is 2.2× the incumbent's — automatic failover across providers is what keeps a bad week from being an outage rather than a datapoint.

A two-column comparison card headed 'Gemini 4 Carbon vs GPT-6 Sol', with rows reading: Intelligence Index 60 versus 44; cost per index task $5.48 versus $2.47; time per task 33.2 versus 7.3 minutes; output speed 62 versus 123 tokens per second; maximum output 65,536 versus 128,000 tokens; input price $2.00 rising to $4.00 on 1 December versus $2.00 flat, with a long-context step to $4.00 above 272,000 tokens. A footer reads 'Index, cost and speed per Artificial Analysis Intelligence Index v5.0.1; prices per Google and OpenAI.'

What to do this week

Three decisions, in the order they resolve.

If the workload is short-answer and high-volume, the question is not Carbon versus Sol at all. It is Sol versus GPT-6.1 Sol, and the successor wins on score at the same price. Carbon's 60 does not enter the argument, because Carbon at 205 million output tokens per index run is the wrong shape for that traffic whatever its score.

If the workload is long-document and the documents vary in length, Carbon is the better buy on 30 November and a worse one on 1 December, and the crossover is arithmetic rather than taste. Model the bill at $4.00 and $20.00 before committing, because that is the rate the contract will run at for most of its life.

If the workload needs audio or video input, there is no comparison to make. Sol cannot take the input, so the choice is Carbon or nothing in this pairing — and the 65,536-token output ceiling becomes the constraint to design around rather than the price.

The one thing not to do is quote the sixteen-point index lead as a reason to switch without the cost-per-task figure beside it. On the same evaluator, same revision, same day, the lead costs 2.2× per completed task and 4.5× the wall-clock time. That is a defensible trade for frontier reasoning and an indefensible one for a queue.

Screenshot of the OrcaRouter model page for openai/gpt-6-sol, showing a 1,050,000-token context window, a 128,000-token maximum output, text and image input with text output, configurable reasoning effort, and pricing of $2.00 per million input tokens and $10.00 per million output tokens with a long-context tier above 272,000 input tokens.

Compared in this article2

Detected from this article · Benchmarks: Artificial Analysis · updated daily