GPT-6.1 Sol: Benchmarks, Pricing, API, and the New Routing Math

Rohit Ramachandran

Sep 29, 2026Updated Sep 29, 2026

• openai
• gpt-6-1-sol
• ai-models
• api-pricing
• benchmarks

GPT-6.1 Sol: Benchmarks, Pricing, API, and the New Routing Math

Seven days ago, the practical OpenAI routing ladder looked straightforward: use GPT-6 Luna to try cheaply, use GPT-6 Sol when acceptance mattered, and reserve GPT-6 Astra for the expensive edge where failure cost more than inference.

GPT-6.1 Sol does not destroy that ladder. It changes the point at which a task should climb it.

OpenAI's DevDay model keeps Sol's familiar $2 per million input tokens and $10 per million output tokens for standard requests up to 272K input tokens. The striking change is not the headline tariff. Cached input falls from GPT-6 Sol's $0.20 to $0.10 per million tokens, only five percent of fresh input. OpenAI also claims near-Astra performance on several agentic evaluations at roughly one-fifth of Astra's task cost.

The launch evidence is OpenAI's evidence. The long-context price cliff still reprices the whole request. Tool-heavy agents can spend more on retries and outputs than they save on input. Ultrafast is live for Astra, not yet priced for GPT-6.1 Sol. And the new model is in the API, ChatGPT Work, and Codex, but was not an ordinary Chat model at the September 29 launch cutoff.

The production question is which routes GPT-6.1 Sol has now earned the right to contest.

For the rest of the launch, see RohitAI's OpenAI DevDay 2026 guide; this analysis stays with Sol's routing economics.

What launched, exactly?

OpenAI announced GPT-6.1 Sol on September 29, 2026 at DevDay. The launch post, model reference, API changelog, and pricing page establish the usable contract.

The API model ID is gpt-6.1-sol. It has a 1,050,000-token context window, 128,000-token maximum output, and an April 30, 2026 knowledge cutoff. It accepts text and image input and produces text. Audio and video are not supported.

Reasoning effort can be set to low, medium, high, xhigh, or max; medium is the default. Unlike some smaller routes, there is no none or minimal setting. Streaming, function calling, and structured outputs are supported. Fine-tuning is not.

The supported tool surface includes web search, file search, image generation, Code Interpreter, hosted shell, apply_patch, skills, computer use, MCP, and tool search. The important contract detail is that tool calling requires the Responses API. Chat Completions can use the model without that tool plane.

At launch, GPT-6.1 Sol was available in the API and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. It was not yet presented as a normal Chat model. US and EU data residency are documented for standard service, while Fast is not available in the EU.

The pricing looks familiar until the cache starts working

For standard requests with no more than 272K input tokens, GPT-6.1 Sol costs per million tokens:

fresh input     $2.00
cached input    $0.10
cache write     $2.50
output         $10.00

That $2/$10 headline matches GPT-6 Sol from the prior week. The cache read does not. GPT-6 Sol launched at $0.20 per million cached tokens; GPT-6.1 Sol halves it to $0.10.

That matters inside an agent repeatedly reading the same repository map, policy corpus, instructions, schemas, or research packet. A cached token costs one-twentieth of a fresh input token.

A cache example that is worth doing on paper

Consider one standard short-context run with 200K tokens of reusable project context, 30K fresh input, and 10K output.

With a warm cache:

200K cached input × $0.10/M = $0.02
 30K fresh input  × $2.00/M = $0.06
 10K output       × $10.00/M = $0.10
total                            $0.18

If all 230K input tokens were fresh, the same arithmetic would be $0.56. That is a 68 percent difference before tools and retries. The first cache-writing call is more expensive, so the gain depends on reuse. The point is not that every request becomes cheap. It is that prefix stability becomes an economic feature.

Teams should version and reuse the large stable prefix, then append changing task state. Repacking the same repository or evidence bundle on every turn can erase one of the model's strongest price advantages.

The 272K cliff can overpower a small quality gain

The model reference applies long-context rates when input exceeds 272K tokens. At that point input and cache categories double, while output rises by 50 percent. The full request is repriced.

A simple comparison shows why this belongs in preflight logic:

270K fresh input + 30K output = $0.54 + $0.30 = $0.84
280K fresh input + 30K output = $1.12 + $0.45 = $1.57

Ten thousand extra input tokens make this illustrative bill 87 percent larger because the request crosses the tier boundary. These are list-price calculations, not predicted invoices; actual cache use, reasoning output, tool fees, and retries change the bill.

That makes retrieval and compaction discontinuously valuable. A context manager that keeps a request just under the threshold may save more than a model-router tweak. But blindly truncating context can lower acceptance enough to lose the saving. The right policy is an eval-backed choice among full context, retrieval, summary, or multiple smaller passes.

How to read OpenAI's benchmark claims

OpenAI's launch case is not based on one academic score. It emphasizes end-to-end software engineering, professional work, automation, computer use, scientific terminal tasks, and factuality. That is the right category of evidence for an agent model because an agent can be expensive even when its tokens are cheap.

But the source is still the model vendor. At the September 29 cutoff, RohitAI did not have a matched independent evaluation that justified declaring GPT-6.1 Sol the universal winner. Harnesses, reasoning settings, fallback rules, task sets, and cost accounting can all move an agent benchmark.

Three cautions keep these numbers useful.

First, “task cost” includes the harness's tokens and tool steps. It is more relevant than token price but less portable to your system.

Second, “near Astra” does not mean identical. A small success-rate difference can dominate economics in a workflow where failure triggers an hour of expert review or a costly external action.

Third, a fallback benchmark evaluates a system, not an isolated checkpoint. That can be good product design, but it must be named.

The new routing math is outcome-shaped

The cheapest route is not necessarily the lowest per-token model. The strongest route is not necessarily the most economical. A useful router estimates this instead:

expected task cost =
  inference + tools + retries + review + failure recovery

expected value per route =
  accepted outcomes × value of success - expected task cost

That formula exposes GPT-6.1 Sol's real opportunity. It can win between Luna and Astra if it eliminates enough retries or review to justify more spend than Luna while preserving most of Astra's completion quality at much lower cost.

Cheap cached input also changes which stage should use Sol. In a long-running agent, the planner or reviewer often rereads a large stable state. That can make Sol attractive for repeated high-judgment checkpoints even when a cheaper model performs bulk execution. The best route may therefore be Sol repeatedly reviewing cached state, not Sol doing every token of work.

Ultrafast is a separate purchasing decision

DevDay also introduced Ultrafast service. In the API, the documented launch call uses GPT-6 Astra with service_tier: "ultrafast" in the Responses API.

OpenAI's recap describes up to six times API speed and up to eight times speed or roughly 300 tokens per second in Codex. The general API guide says up to eight times standard. Those statements are not identical, so production planning should use measured latency for a specific route rather than a single keynote multiplier.

The list price makes the trade clear. For Astra at short context, Ultrafast costs $60 input, $6 cached input, $75 cache write, and $300 output per million tokens. At long context it costs $120, $12, $150, and $450. That is six times Astra's standard list rate.

In other words, the claimed acceleration and price premium are of roughly the same order. Ultrafast is not a free optimization. It is a way to buy wall-clock time when latency has direct business value.

Ultrafast was broadly available through the API at low default limits: 500K tokens per minute for tiers 1 through 3, 1M for tier 4, and 5M for tier 5. It supports Global and US residency, not EU or other regional residency. OpenAI recommends WebSockets for tool-heavy workloads to reduce repeated connection overhead.

GPT-6.1 Sol Ultrafast was announced as coming soon. No Sol Ultrafast price was listed at the cutoff. Do not apply Astra's price or promise Sol's availability by analogy.

All figures are dollars per million tokens. Fast service is separately priced at two times the standard lane. Batch and Flex are 50 percent below standard, while regional processing adds 10 percent according to the pricing documentation. Confirm the current table before deployment because service tiers are particularly likely to change.

Pro 500 bundles speed, not proof of value

The new ChatGPT Pro tiers documentation lists Pro 500 at $500 per month. OpenAI's recap says it carries 25 times the Plus allowance, and the Help Center says it includes Astra Ultrafast. Pro 100 and Pro 200 do not include Ultrafast at launch, and buying credits on those plans does not unlock it.

This tier fits heavy operators whose bottleneck is waiting for Work or Codex. “Included” does not remove opportunity cost, rate limits, or the need to route.

Evaluate Pro 500 with time saved per accepted task:

monthly value =
  accepted tasks accelerated
  × minutes saved per task
  × value of operator time
  - added subscription cost

Do not justify it with tokens you could theoretically consume. Justify it with blocked work that actually becomes faster.

A clean API starting point

For a new integration, use the Responses API and make reasoning effort explicit. This minimal JavaScript example intentionally leaves tools out so the first comparison isolates model behavior:

import OpenAI from 'openai';

const client = new OpenAI();

const response = await client.responses.create({
  model: 'gpt-6.1-sol',
  reasoning: { effort: 'medium' },
  input: 'Review this migration plan. Return risks, missing tests, and a go/no-go recommendation.'
});

console.log(response.output_text);

Then add production tools, schemas, and approvals in a second eval. Capture response IDs, model and service tier, token usage, tool calls, retries, elapsed time, approval events, and the final acceptance decision.

The seven-day migration plan

Do not replace GPT-6 Sol in one deployment. Run a short controlled migration that keeps the old route available for replay.

A good result may be a partial migration. Sol 6.1 could win repository review and document synthesis while Luna keeps extraction and Astra keeps a narrow set of costly edge cases. That is not indecision. It is routing.

What could change after launch

Several details were still moving at the research cutoff.

GPT-6.1 Sol Ultrafast was announced but not live or priced. Independent labs had not yet supplied a broad matched verdict. Product rollouts can lag documentation by account or region. OpenAI's API changelog labels multi-agent support beta. Ordinary Chat availability was not announced. And session recordings from DevDay were still pending while this analysis was prepared.

Those are reasons to timestamp the conclusion, not to ignore the release.

As of September 29, the evidence supports three actions: benchmark GPT-6.1 Sol seriously, redesign stable context for cache reuse, and make the 272K threshold visible in routing. It does not support deleting every other model route.

FAQ

How much does GPT-6.1 Sol cost?

For standard API requests up to 272K input tokens, OpenAI lists $2 per million fresh input tokens, $0.10 cached input, $2.50 cache writes, and $10 output. Above 272K input, the whole request uses $4, $0.20, $5, and $15 rates respectively.

Is GPT-6.1 Sol cheaper than GPT-6 Sol?

Fresh input and output launch rates are the same at $2/$10 for standard short-context use. The major direct tariff improvement is cached input, which falls from $0.20 to $0.10 per million tokens. Real task cost can also change because the models may consume different tokens, steps, and retries.

Is GPT-6.1 Sol as good as GPT-6 Astra?

OpenAI reports near-Astra results on several evaluations at substantially lower task cost, but not parity everywhere. Astra remains ahead in at least some launch results, including Terminal-Bench Science 0.1. There was no independent universal verdict at the research cutoff.

Does GPT-6.1 Sol support tools and computer use?

Yes. Its documented tool surface includes computer use, web and file search, code execution, hosted shell, MCP, skills, and other tools. Tool calling requires the Responses API.

What happens above 272K tokens?

The threshold is based on input tokens. Once input exceeds 272K, the full request moves to the long-context rates: input and cache categories double and output becomes 1.5 times the short-context rate. It is not marginal pricing only on the excess.

Is GPT-6.1 Sol Ultrafast available?

Not at the September 29 launch cutoff. OpenAI said it was coming soon. Astra Ultrafast was live, with separate pricing and Global/US residency only.

Should I buy Pro 500 for GPT-6.1 Sol?

Only if the higher allowance and included Astra Ultrafast save enough time on accepted Work or Codex tasks to justify $500 per month. Pro 500 economics are subscription and workflow economics, not the same as API token pricing.

Final take

GPT-6.1 Sol is a more interesting release than its unchanged $2/$10 headline suggests.

It makes stable context cheaper to reread. It makes Astra-like outcomes plausible for more workflows. It pushes task cost, not token cost, into the center of the model decision. And it arrives beside an acceleration lane expensive enough to force an explicit value-of-time calculation.

The best migration is therefore not a global model-name replacement.

Put GPT-6.1 Sol into the routes where judgment repeats over large, reusable context. Keep Luna on cheap, bounded work. Keep Astra only where it wins a meaningful edge. Add human authority where the action is irreversible. Count the request before the 272K cliff. Measure review and retries after it runs.

Seven days ago, “build for acceptance” was the right advice. GPT-6.1 Sol does not overturn it.

It makes the acceptance calculation more favorable—and more demanding.