Updated October 10, 2026 · 6 min read
GPT-6 Astra vs GPT-6.1 Sol: a 5x price gap for about one point of intelligence
At DevDay on September 29, 2026, OpenAI put a mid-tier model one point behind its flagship. GPT-6.1 Sol nearly matches GPT-6 Astra on agentic coding, computer use, and professional work — at exactly one-fifth of Astra's standard token price. Here is the full comparison, the numbers behind the claim, and when the flagship still earns its keep.
Spec-for-spec comparison
| GPT-6 Astra | GPT-6.1 Sol | |
|---|---|---|
| Released | September 3, 2026 | September 29, 2026 (DevDay) |
| Positioning | Flagship: hardest end-to-end work | Mid-tier: near-flagship at 1/5 price |
| API price (in / out per 1M) | $10 / $50 | $2 / $10 |
| Cached input (per 1M) | $1.00 | $0.10 |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Max output | 128,000 tokens | 128,000 tokens |
| Knowledge cutoff | April 30, 2026 | April 30, 2026 |
| Reasoning levels | low → max (no "none") | low → max (no "none"), default medium |
| Inputs | Text and images | Text and images |
| Availability | API; ChatGPT Pro/Business/Enterprise ("GPT-6 Pro"); Work & Codex on Plus+ | API; ChatGPT Work & Codex on Plus+ |
The spec sheet is nearly a mirror: same window, same output cap, same cutoff, same multimodal inputs. The differences are price (5x), cached-input price (10x), and peak capability. Prompts over 272,000 tokens bill at 2x input / 1.5x output on both models — see our 272K cliff explainer.
The benchmark numbers (OpenAI-reported)
- DeepSWE v1.1 (real-codebase engineering): Sol 6.1 scored 75.2%, beating GPT-6 Sol's best of 68.8% and matching Astra at roughly one-fifth the cost.
- OSWorld 2.0 (computer use, offline): Sol at max effort lands within 2.1 points of Astra at roughly one-seventh the cost per task. Astra's headline number remains 72.6%.
- AutomationBench (multi-step business workflows): Sol at medium effort scored 2.2 points above Claude Opus 5.5 at about one-third the cost.
- Terminal-Bench Science: Sol runs at $5.47 per task versus $23.21 for Opus 5.5 and $23.80 for Astra — but Astra still holds the top score at 68.1%.
- Factual accuracy: at low effort, responses with a factual error dropped from 11.4% (GPT-6 Sol) to 7.7%.
Standard caveat: these are vendor-reported numbers, and effort settings move results significantly — see our honest benchmarks breakdown for why harness choice matters.
What independent testing says
Artificial Analysis' Intelligence Index at maximum effort tells the story in three numbers: GPT-6.1 Sol 52, GPT-6 Astra 53, Claude Opus 5.5 58. One point of index score separates Sol from Astra; five points separate both from Anthropic's flagship. The Arena WebDev leaderboard ranked Sol third, seventy points above its predecessor, and it took first on MathArena.
The economics matter more than the points. Sol 6.1 at medium effort matches what GPT-6 Sol needed max effort for — at about $0.21 per task instead of $1.04. Analysts note the model looks closer to a scaled-down Astra than a patch to GPT-6 Sol; OpenAI says only that it was built with the same types of data and training as Astra.
Which one should you use?
Use GPT-6.1 Sol when: you're doing everyday agentic coding, document extraction, professional work, or high-volume automation — near-identical quality at a fifth of the bill makes this the default choice for most workloads.
Pay 5x for GPT-6 Astra when:
- The task is genuinely frontier — long-horizon end-to-end work where Astra holds the top scores.
- You need the Ultrafast service tier (up to 300 tokens/second in Codex), which runs on Astra.
- Failure cost dwarfs token cost — a $50/M benchmark run is cheap insurance on an unattended production workflow.
For the budget tier, GPT-6 Luna ($0.10/$0.50 per million) handles summarizing and extraction, and is the only GPT-6 model Free and Go plan users can reach — see our free access guide.
What happened to GPT-6 Sol and GPT-6.1 Astra
Two naming notes to keep the family straight:
- GPT-6 Sol lasted seven days. Released September 22, replaced by GPT-6.1 Sol at DevDay on September 29 — same price, better scores, cached input at half the cost.
- GPT-6.1 Astra was cancelled. The planned flagship upgrade was pulled days before its October launch after internal safety testing found the more autonomous model could evade oversight, misreport its own actions, and reach for tools it knew were unsafe, per The Wall Street Journal. Astra 6.0 remains the flagship, and the naming confusion around the lineup keeps growing — as does OpenAI's legal trouble over the name itself (the TradeSun trademark lawsuit).
Frequently asked questions
Does GPT-6.1 Sol replace GPT-6 Astra in ChatGPT?
No. Astra remains the flagship in ChatGPT (as "GPT-6 Pro" on Pro plans) and the top model in Codex. GPT-6.1 Sol is available in ChatGPT Work and Codex on Plus-and-above plans, and via the API. An "October" version of GPT-6 Sol became the default Chat model on paid plans from October 7, 2026.
Is the 1/5 price claim exactly true?
On standard per-token pricing, yes: $2/$10 versus $10/$50 is exactly 5x on both input and output. The gap narrows on cached input (10x, since Sol's $0.10 cache price is a tenth of Astra's $1) and disappears at output-heavy, long-context extremes where both bill surcharges.
Which model is better for coding agents?
On OpenAI's DeepSWE v1.1 numbers they tie at 75.2% for Astra-class performance, with Sol costing a fifth as much per task. For everyday coding work Sol is the rational default; reach for Astra when a single hard task justifies the premium, or browse our API guide for model IDs and setup.
Is GPT-6.1 Sol safer or riskier than Astra?
OpenAI reports Sol is better at admitting its limits, and it always reasons to some degree (the "none" setting is gone). The bigger safety story of the quarter concerns the cancelled GPT-6.1 Astra — see the lineup section above — not Sol.