GPT-6.1 Sol vs GPT-6 Sol: smarter but slower

OpenAI released GPT-6 Sol on September 22, 2026. Seven days later, it replaced it with GPT-6.1 Sol: same price, same context window, and a much stronger model. Few teams had time to finish evaluating the first release before the second arrived.

The new model is smarter on nearly every test we track. In our Ready for Enterprise (RFE) LLM Benchmark, its capability score jumps from 70.3 to 88.6. However, it also writes much more slowly, at 29 tokens per second versus 69 for GPT-6 Sol. That slowdown is large enough that GPT-6 Sol still edges ahead on our overall enterprise score, 58.1 to 57.3.

So the real question isn't which model is better, but how long your responses are, and whether anyone is waiting for them.

What changed in one week

Sources: OpenAI (GPT-6 Sol), OpenAI (GPT-6.1 Sol), The Batch, and the FullStack RFE LLM Benchmark.

OpenAI describes GPT-6.1 Sol as near-Astra intelligence for a fifth of the price. GPT-6 Astra is OpenAI's most intelligent model, and GPT-6.1 Sol approaches its results on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices. GPT-6.1 Sol is available through the OpenAI API and, at launch, in ChatGPT Work and Codex for paid plans, though not yet in regular ChatGPT conversations.

The full benchmark comparison

Source: FullStack RFE LLM Benchmark, data snapshot used for this article. Rankings can change as new models ship and results are added.

Our RFE score combines six public benchmarks covering agentic work, factual accuracy, human preference, web development, graduate-level science, and advanced math, then adds price and speed. The capability score reflects benchmark performance alone. GPT-6.1 Sol wins every capability row in the table, yet its RFE score is slightly lower because the speed penalty outweighs its gains.

Where the extra intelligence shows up

The biggest gains are in agentic and professional work, the tasks where a model has to plan, use tools, and stay accurate over many steps.

Complex software engineering

On Terminal-Bench 4.0, GPT-6.1 Sol scores 58.2% versus 49.4% for GPT-6 Sol. On DeepSWE v1.1, which evaluates complex software engineering tasks in real codebases, OpenAI reports that GPT-6.1 Sol beats GPT-6 Sol's best score by 6.4 percentage points, and does so at a lower reasoning effort and cost. It also matches GPT-6 Astra on that test at roughly one-fifth of the cost. On Arena WebDev, its rating rose from 1683 to 1753.

Multi-step business workflows

AutomationBench tests whether agents complete end-to-end workflows using 47 tools across sales, marketing, operations, support, finance, and HR. At medium reasoning effort, GPT-6.1 Sol scores 4.8 percentage points higher than GPT-6 Sol at the same setting.

Computer use

On OSWorld 2.0's offline set, which OpenAI reports as a partial reward score, GPT-6.1 Sol outperforms GPT-6 Sol by seven percentage points at maximum reasoning effort, at less than half the cost. It comes within 2.1 percentage points of GPT-6 Astra, the more powerful model, at roughly one seventh the cost per task.

Scientific research

The largest single jump is on Terminal-Bench Science, which tests scientific research workflows such as data analysis, simulation, and theorem proving. At maximum reasoning effort, GPT-6.1 Sol more than doubles GPT-6 Sol's score at less than half the cost per task. OpenAI still recommends GPT-6 Astra for the most difficult scientific research.

Factual accuracy

On SimpleQA Verified, GPT-6.1 Sol scores 73.9%, up from 60.7%. OpenAI's own evaluations point the same way: at low reasoning effort, the share of responses containing a factual error fell from 11.4% to 7.7% on a set of deliberately difficult prompts. The new model is also more candid about its limits. In one test, it failed to tell users their search tool was broken in 2.1% of attempts, compared with 4.9% for GPT-6 Sol.

Where the gains are smaller

On graduate-level science and math, GPT-6 Sol was already strong, and the two models show similar performance. GPQA Diamond rose only from 94.3% to 95.4%, and FrontierMath Tiers 1–3 from 89.8% to 93.7%. If your workload is mostly this kind of reasoning, the upgrade matters less.

Independent testing confirms the overall gain. On the Artificial Analysis Intelligence Index, a composite of 10 benchmarks, GPT-6.1 Sol at max reasoning scores 52, four points above GPT-6 Sol's 48 and one point below GPT-6 Astra.

Where the speed goes

In our measurements, GPT-6 Sol generates output about 2.4 times faster than GPT-6.1 Sol. Artificial Analysis measured a smaller gap, at 57.7 tokens per second for GPT-6.1 Sol versus 87.6 for GPT-6 Sol. Exact numbers vary by provider, load, and settings, but both measurements show the new model running noticeably slower.

The delay isn't at the start. Time to first token is nearly identical: 3.51 seconds for GPT-6.1 Sol and 3.60 for GPT-6 Sol. The gap opens once the model starts writing, so it grows with every token of output. Here's what that means for responses of different lengths, using our measured averages:

Estimates use time to first token plus output length divided by measured output speed. Real response times also depend on reasoning effort, tool calls, and load.

For a short answer, users may barely notice the difference. For a long report, the wait more than doubles. For an agent that makes dozens of sequential calls, those extra seconds add up across the whole run.

OpenAI has announced GPT-6.1 Sol Ultrafast, with up to 8x faster token generation compared with its standard speed in Codex. If speed is what's holding you back on GPT-6 Sol, check whether Ultrafast is available for your use case, and at what cost, before you decide.

Which model for which job

For latency-critical, user-facing apps, test GPT-6.1 Sol on your real traffic and response lengths before switching. Also check OpenAI's model documentation for GPT-6 Sol's availability timeline before you plan around staying on it.

Pricing and caching

Both models cost $2 per million input tokens and $10 per million output tokens, for a FullStack blended price of $4.00. The one price change favors the new model: GPT-6.1 Sol charges $0.10 per million cached input tokens, 95% less than standard input and half of GPT-6 Sol's cached rate. Agents that reuse long system prompts or shared context across requests benefit most.

Both models have a 1,050,000-token context window with up to 128,000 output tokens. GPT-6.1 Sol's standard rates apply to prompts up to 272,000 input tokens, with higher rates above that, so check OpenAI's current pricing before sending very large documents in a single request.

OpenAI’s results also show GPT-6.1 Sol often reaching better scores at lower total cost per task. Compare the cost of a completed task, not just the per-token price.

Before switching: reasoning settings

The biggest code change is reasoning control. GPT-6.1 Sol supports five reasoning settings from low to max, with medium as the default, and it removes GPT-6 Sol's option to turn reasoning off. If you run GPT-6 Sol with reasoning off to save time, switching means choosing a reasoning level and retesting both latency and cost.

A practical way to evaluate the switch:

1. Record a GPT-6 Sol baseline: prompts, tools, reasoning settings, typical response lengths, and acceptance criteria.
2. Update any requests that turn reasoning off, and set the reasoning level explicitly.
3. Run both models on the same requests and compare completion rate, factual errors, total response time, and cost per accepted result.
4. Move quality-first workloads first, and test latency-sensitive features separately before deployment.

Many teams will run both models, sending accuracy-critical work to GPT-6.1 Sol and speed-critical work to GPT-6 Sol. Our AI Product Development services help teams build that kind of routing into production applications.

How FullStack can help

FullStack builds OpenAI-powered systems for clients and runs five of its own in production, covering retrieval, prospecting, voice, client-facing search, and talent matching. Our OpenAI implementation services help enterprises evaluate new model versions on their real workloads, migrate without surprises, and route each task to the model that handles it best.

^{Compare these models in the RFE benchmark.}