# GPT-6.1 Sol Is in the API: Near-Astra at One-Fifth the Cost > Source: > Published: 2026-10-06 11:08:39+00:00 OpenAI shipped GPT-6.1 Sol on September 29 — the same day as DevDay — and it quietly rewrites the cost math for most production AI workflows. The model scores 75.2% on DeepSWE v1.1, beating Astra’s 74.1% on that coding benchmark, at $2/$10 per million input/output tokens versus Astra’s $10/$50. Prompt cache is now $0.10 per million — half of what GPT-6 Sol charged a week earlier. If you’re running agentic coding loops on Astra today, there’s arithmetic to do. ## The Pricing Reality The headline number is 5x cheaper than Astra on standard input and output. What doesn’t make the headline: cached input just halved again. GPT-6 Sol was $0.20 per million cached tokens. GPT-6.1 Sol is $0.10. Astra is $1.00. For agent loops that pass large context repeatedly — which is most of them — that’s a 10x cache advantage over Astra and a meaningful additional cut from last week. | Model | Input | Cached Input | Output | |---|---|---|---| | GPT-6 Astra | $10.00 | $1.00 | $50.00 | | GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | | GPT-6 Luna | $0.10 | — | $0.50 | [OpenAI pricing page](https://developers.openai.com/api/docs/pricing). On the [DeepSWE v1.1 benchmark](https://artificialanalysis.ai/articles/gpt-6-1-sol-replaces-gpt-6-sol-after-just-7-days-with-near-astra-intelligence), which runs real-world software engineering tasks across live codebases, the per-task cost drops from $4.43 on Astra to $0.65 on GPT-6.1 Sol — 6.8x cheaper per completed task. The model also supports Batch and Flex processing modes at 50% off standard rates for async workflows. ## What "Near-Astra" Actually Means OpenAI’s marketing says "near-Astra performance." On coding specifically, that’s underselling it: GPT-6.1 Sol scores 75.2% versus Astra’s 74.1% on DeepSWE v1.1. On computer use and broader agentic tasks (OSWorld 2.0), the gap widens to 2.8 percentage points — 54.5% for Sol versus 57.3% for Astra. That gap is real but narrow, and at one-fifth the cost, it rarely justifies using Astra for routine agent work. Compared to its predecessor, GPT-6 Sol, this upgrade is significant: +6.4 points on DeepSWE, +7 on OSWorld, and a factual error rate of 7.7% versus 11.4%. OpenAI pushed this within seven days of GPT-6 Sol’s launch — which is either a sign of rapid iteration or an admission that the original Sol was underbaked. ## Three Migration Catches Swapping `gpt-6-sol` for `gpt-6.1-sol` works for most setups. Watch for these three before deploying: **Reasoning effort values changed.** GPT-6.1 Sol dropped support for `reasoning_effort: "none"` and `"minimal"`. The minimum is now `"low"`. This fails silently in some implementations — audit your code before switching. **Tool calling requires the Responses API.** If your application uses function calling through the Chat Completions endpoint, you need to migrate to the Responses API first. GPT-6.1 Sol does not support tool calling on Chat Completions. **Latency increases.** Time to first token is 63% slower than GPT-6 Sol; output speed is 13% slower. More reasoning depth per token produces better results but adds latency. For interactive applications, benchmark before switching in production. ## When to Keep Astra For most agentic coding workloads, GPT-6.1 Sol is the new default. Keep Astra for three scenarios: one-shot creative coding (early community feedback puts Sol behind competitors on UI generation and game scaffolding); latency-critical paths where Astra Ultrafast is already live at 300 tokens per second (Sol Ultrafast is still "coming soon"); and edge cases where that 2.8-point computer use gap in production metrics matters more than a 5x cost reduction. ## What to Do Now The model is live as `gpt-6.1-sol` across [the API and ChatGPT Work](https://openai.com/index/introducing-gpt-6-1-sol/). The switch is a one-line model ID change for most setups. Check your `reasoning_effort` values, confirm you’re on the Responses API if you use tool calling, and run a cost estimate against your current Astra spend. Read the full [DevDay 2026 recap](https://www.infoq.com/news/2026/10/openai-devday-2026/) for the rest of what shipped. The math on Sol is hard to argue with for coding-heavy workloads.