OpenAI introduced GPT-6.1 Sol at DevDay — an upgrade to GPT-6 Sol that the company says nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth of Astra’s standard API input and output token prices. That is the headline for builders: Astra-class work loops without Astra-class bills, with a sharper cache story for agents that reuse context.
GPT-6.1 Astra is not shipping. Reporting earlier this week said OpenAI shelved that upgrade over safety/alignment concerns. DevDay still moved the mid-tier: Sol 6.1 is live in ChatGPT Work and Codex for Plus through Edu, and on the API as gpt-6.1-sol. Same-day, OpenAI also launched Dots (always-on agents) — useful context, not the lead of this piece.
Image credit: OpenAI
Key points
- What shipped: GPT-6.1 Sol (
gpt-6.1-sol) — upgrade over GPT-6 Sol; OpenAI’s claim is near-Astra intelligence on coding, computer use, and professional work at 1/5 Astra’s standard input/output rates. - API pricing (per 1M tokens, standard short context): Input $2 · Cached input $0.10 · Cache writes $2.50 · Output $10. Astra for comparison: $10 / $1 / $12.50 / $50. Cached input is 95% below Sol 6.1’s own uncached input and 50% below GPT-6 Sol’s $0.20 cache reads.
- Where: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Not yet in Chat. API via Responses (tools) and Chat Completions (no tool calling). Ultrafast (up to 8× faster token generation in Codex) promised in the coming days.
- Shape: ~1.05M context · 128K max output · knowledge cutoff April 2026 · reasoning effort
low/medium(default) /high/xhigh/max(noneandminimalunsupported). - Vendor benches (self-reported; effort-/harness-coupled): DeepSWE v1.1 matches Astra and beats GPT-6 Sol by 6.4 points at lower effort/cost · GDP.pdf above Opus 5.5 with fallbacks at under half the cost per task · AutomationBench +2.2 points vs Opus 5.5 at medium (~1/3 cost) · OSWorld 2.0 offline within 2.1 points of Astra at max (~1/7 cost/task) · Terminal-Bench Science ~$5.47/task vs Opus ~$23.21 / Astra ~$23.80.
- Factuality: At low effort, share of responses with a factual error drops from 11.4% (GPT-6 Sol) to 7.7%; across tested settings, within 1.9 points of Astra — on deliberately hard, error-flagged prompts, not typical chat.
- Safety posture: Treated as Critical cybersecurity and High biological/chemical under OpenAI’s Preparedness Framework; same safeguards stack as Astra. Alignment evals improve vs GPT-6 Sol (transparency, restrictions, unauthorized outcomes); no observed automated-reviewer bypass attempts in the reported tests.
- Same-day DevDay context: Dots — always-on Astra-powered agents for Pro / Business Premium / Enterprise; Slack/Teams; cloud computer + 4,000+ apps. Mention only — not the lead.
- Caveats: All peer scores and $/task figures are vendor-reported. 9to6AI has not hands-on audited coding quality or agent success rates. Re-measure on your harness before flipping production defaults.
What shipped
OpenAI positions GPT-6.1 Sol as the cost-efficient workhorse next to Astra: keep Astra for the hardest scientific and judgment-heavy jobs; use Sol 6.1 when you want most of that agentic coding and computer-use ability without the Astra sticker.
| Field | Value |
|---|---|
| Model id | gpt-6.1-sol |
| Family | GPT-6.1 mid-tier (Astra non-ship for 6.1; Sol/Luna line continues) |
| Surfaces | ChatGPT Work, Codex, OpenAI API |
| Standard pricing | $2 in / $0.10 cache read / $2.50 cache write / $10 out per 1M |
| Context / cutoff | ~1.05M context; 128K max output; knowledge cutoff April 2026 |
| Reasoning effort | low · medium (default) · high · xhigh · max |
| Not yet | Chat surface; Ultrafast shipping “coming days” |
Official announcement: openai.com/index/introducing-gpt-6-1-sol.
What changed vs GPT-6 Sol, Astra, and the market
Same Sol list price, sharper cache. Standard $2/$10 matches GPT-6 Sol’s short-context rates, but cached input drops from $0.20 to $0.10 — the number that matters for agent loops that pin a large system prompt or tool schema across turns. Versus Astra’s $10/$50, Sol 6.1 is literally one-fifth on both input and output.
Vendor coding and workflow tables. On DeepSWE v1.1, OpenAI says Sol 6.1 matches Astra while beating GPT-6 Sol by 6.4 points at lower effort and cost. GDP.pdf (complex professional PDFs) is framed above Opus 5.5 with fallbacks at under half the cost per task, approaching Astra at roughly one-fifth. AutomationBench (47-tool multi-step business workflows) is +2.2 points vs Opus 5.5 at medium effort and about a third of the cost. Treat every peer row as directional until independent harnesses settle.
Computer use and science. OSWorld 2.0 offline: +7 points over GPT-6 Sol at max effort, within 2.1 points of Astra at roughly one-seventh the cost per task. Terminal-Bench Science: more than doubles GPT-6 Sol at max; OpenAI quotes ~$5.47 per task vs ~$23 for Opus 5.5 and Astra — while still saying Astra leads the science scoreboard (68.1%) for the hardest research workflows.
Factuality and alignment. The biggest factuality jump vs GPT-6 Sol is at low effort (11.4% → 7.7% error-containing answers on hard, previously flagged prompts). Alignment evals show lower failure rates than GPT-6 Sol on broken-search transparency, respecting restrictions, and avoiding unauthorized agent outcomes — closer to Astra’s posture, with the same “no reviewer bypass observed” note as Sol and Astra in the launch materials.
What did not ship. GPT-6.1 Astra remains shelved after internal safety tests. That makes Sol 6.1 the practical DevDay upgrade for teams that were waiting on a 6.1 mid-tier, not a substitute for a new frontier Astra drop.
Same-day Dots (brief). OpenAI also launched Dots: always-on agents on Astra with a cloud computer, app connectors, Slack/Teams, and custom permission rules, rolling out to Pro / Business Premium / Enterprise. Separate product story — cover it on its own if you need the Muse-competitor angle; builders picking a default model should start with Sol 6.1 pricing and eval fit.
What it means for builders
If you are on GPT-6 Sol today. Plan a move to gpt-6.1-sol for agentic coding, computer-use loops, and PDF-heavy professional workflows. Re-measure tokens/task and wall-clock — list price is unchanged, but cache reads are half of Sol’s, and vendor tables claim higher success at similar or lower effort.
If you are on GPT-6 Astra for cost reasons. Keep Astra for the hardest science and judgment jobs OpenAI still owns on the vendor tables. Route high-volume coding agents, OSWorld-style computer use, and AutomationBench-like business workflows to Sol 6.1 and compare quality/$ on your own traces before cutting Astra out entirely.
If cache and agent volume dominate spend. $0.10/M cached input is the lever. Pin stable prefixes, tool schemas, and long instructions; model Batch/Flex (50% off standard) and the 2× long-context band above 272K input before you celebrate the sticker.
If you need ChatGPT “Chat” or Ultrafast now. Sol 6.1 is in Work and Codex, not Chat yet. Ultrafast (up to 8× faster generation in Codex) is “coming days” — do not block a migration on a date OpenAI has not pinned.
If you are comparing Anthropic’s mid-tier. Claude Sonnet 5.5 also sits at $2/$10 with a different bench stack and cyber fallback story. Pick by harness fit (DeepSWE / OSWorld / AutomationBench vs Terminal-Bench / CursorBench), not by press-cycle vibes. See 9to6AI’s Sonnet 5.5 explainer.
Safety / compliance teams. Expect Astra-class preparedness labels (Critical cyber, High bio/chem) and the Astra safeguards stack. Read the system card addendum before enabling broad tool use in production.
What to watch
- Independent Artificial Analysis / third-party harness numbers once they settle (DeepSWE, OSWorld, AutomationBench, tokens-per-task vs Sol and Astra).
- Ultrafast Codex rollout and whether Chat surface support lands soon after.
- Real-world agent success vs Astra when cache discounts dominate the bill.
- Whether teams standardize on Astra-for-hardest / Sol-6.1-for-volume as the default GPT-6 stack.
- How Dots’ always-on agents change demand for Sol vs Astra once Pro/Enterprise usage data appears.
