GPT-6.1 Sol vs. GPT-6 Astra: Same accuracy at 18% of the cost

OpenAI launched GPT-6.1 Sol on September 29, just a week after it launched GPT-6 Sol. OpenAI marketed 6.1 as an upgrade to 6 and as a cheaper near-match for Astra. OpenAI says the new model nearly matches Astra on agentic coding, computer use, and professional work. It also says GPT-6.1 Sol makes fewer factual errors than GPT-6 Sol at low reasoning effort. OpenAI calls it “near-Astra intelligence for a fifth of the price.” OpenAI also halved cached-input pricing, from GPT-6 Sol’s $0.20 to $0.10.

GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. GPT-6 Astra costs $10 per million input tokens, $50 per million output tokens, and $1 per million cached input tokens. That puts GPT-6.1 Sol at one-fifth of Astra’s price for input and output tokens, and one-tenth for cached input. 

These claims are a lot for my skeptical self to take in. Does the price gap hold up once real token use is measured? And how close does GPT-6.1 Sol actually come to Astra’s intelligence? To find out, I ran both models through the same three tests I ran on GPT-6 Sol and Opus 5.5. 

The tests

I called both models through the OpenAI Responses API with identical prompts, reasoning effort set to max, and a 64,000-token output limit. These are the same settings I used for GPT-6 Sol. The tests work for GPT-6.1 Sol vs Astra because they each map to one of OpenAI’s marketing claims for GPT-6 Sol and GPT-6.1 Sol.

• CI triage – The model reads 40 failed CI job logs and a runbook with conditional rules. For each job, it decides whether to retry, block, or page.
• Incident logs – The model reads 3,664 lines of logs from five services during a two-hour outage and answers seven postmortem questions. The answers depend on exact counting, a timezone conversion, and a red herring.
• Resolver spec – The model writes a dependency resolver for a fictional package manager from a two-page spec, without running any code. A hidden suite of 120 tests grades the result.

I didn’t include prompts for this because I used multiple files, repos, and lengthy prompts rather than something that copies/pastes easily. 

CI triage

Both models made all 40 calls correctly on all five runs. Astra was faster, averaging 24 seconds per run to GPT-6.1 Sol’s 31 seconds. Both read 3,867 input tokens per run, and output was close, with 1,381 tokens for GPT-6.1 Sol and 1,431 for Astra. GPT-6.1 Sol cost about 2 cents per run, and Astra cost 11 cents.

Accuracy was a tie. Astra won on speed, and GPT-6.1 Sol won on cost.

Incident logs

This test gave GPT-6 Sol the most trouble last month. It missed a customer on two runs and miscounted failed checkouts on another. GPT-6.1 Sol and Astra both answered all seven questions correctly on every run.

GPT-6.1 Sol averaged 2 minutes and 20 seconds per run, about 19% faster than Astra’s 2 minutes and 53 seconds. Each run read 113,966 input tokens. Output was nearly even, at 8,316 tokens for GPT-6.1 Sol and 8,539 for Astra. GPT-6.1 Sol cost $0.31 per run, and Astra cost $1.57.

Accuracy was a tie. GPT-6.1 Sol won on speed and cost.

Resolver spec

Last month, GPT-6 Sol left a stray parenthesis in one run that crashed the resolver. This time, both models passed all 120 hidden tests on every run.

This test showed the biggest gap. GPT-6.1 Sol averaged 7 minutes and 7 seconds per run, about 30% faster than Astra’s 10 minutes and 6 seconds. Astra wrote 25,207 output tokens per run to GPT-6.1 Sol’s 19,637, or 28% more. Counting input and output, that added up to $1.28 per run, compared with $0.20 for GPT-6.1 Sol.

Accuracy was a tie. GPT-6.1 Sol won on speed, tokens, and cost.

Results

Both models were perfect on all 15 runs. GPT-6.1 Sol’s runs took 49 minutes and 54 seconds and cost $2.66 in total. Astra’s took 1 hour, 6 minutes, and 55 seconds and cost $14.77. GPT-6.1 Sol cost 18% of what Astra did, close to the one-fifth price OpenAI advertises. Output tokens were nearly the same on CI triage and incident logs, and Astra used 28% more on the resolver spec.

What do I think?

OpenAI’s claim held up in my tests. GPT-6.1 Sol matched Astra on every run, cost less than a fifth as much, and was faster on the two longer tests. 

I’d use GPT-6.1 Sol for this kind of work. It got every answer right on all three tests, like Astra, for less than a fifth of the price. My tests can’t show the two models are equal on everything, and Artificial Analysis’ index still puts Astra slightly ahead at max effort, but nothing I ran gave me a reason to pay for Astra.