GPT-6.1 Sol: The Budget AGI That Matches Astra at One-Fifth the Cost
The context
At DevDay 2026 on September 29, OpenAI shipped GPT-6.1 Sol alongside its headline launch of Dots. But Sol may be the more consequential release. It ties GPT-6 Astra on DeepSWE v1.1 — the gold-standard benchmark for real-world software engineering — while costing one-fifth as much.
Where Astra charges $10/$50 per million tokens, Sol is priced at $2/$10. That is not a discount. It is a repositioning of the entire frontier tier.
The benchmark numbers
OpenAI reports the following for GPT-6.1 Sol on its five headline evaluations:
| Benchmark | GPT-6.1 Sol | GPT-6 Astra | Gap |
|---|---|---|---|
| DeepSWE v1.1 | 75.2% | 75.2% | Tied |
| OSWorld 2.0 | Within 2.1 pts | Within 2.1 pts | Nearly tied |
| AutomationBench | 2.2 pts above Opus 5.5 | Leader | Trailing Sol |
| Artificial Analysis Index | 51.8 | 61 | Astra leads by 9.2 |
| Terminal-Bench Science 0.1 | Trailing | Leader | Astra leads |
Sol ties or nearly ties Astra on coding and agent benchmarks, but trails clearly on science, computer use, and cyber evaluations. The model is purpose-built for agentic workloads — the reason it exists.
Independent analysis by Emergent confirms: Sol costs $0.72 per index task, versus Astra's $23.80 per Terminal-Bench Science task and $7.63 per benchmark task. That is a 10-33x cost advantage
Why Sol exists
The agentic numbers are the reason this model exists. As the Datacamp guide to GPT-6.1 Sol explains, Sol is designed to be the default model inside Codex — the coding agent environment. Codex CLI 0.159.1 made gpt-6.1-sol the default model, and the Pro 500 plan ($500/month) bundles access to Sol alongside Ultrafast, OpenAI's new paid speed tier.
Sol is not a general-purpose model. It is a specialized agent that prioritizes cost-efficiency for coding workflows over raw capability on science and cyber tasks. This is OpenAI's answer to DeepSeek's V4.1 Flash — the budget-conscious agent that does 95% of what the flagship does at 20% of the price.
The cost-per-task inversion
At $0.72 per task on the Artificial Analysis index, Sol is within striking distance of DeepSeek V4.1 Flash ($0.15/$0.60 off-peak). And unlike DeepSeek, Sol is a closed model — it comes with OpenAI's support, SLA guarantees, and integration with the broader OpenAI ecosystem (Dots, Codex, ChatGPT Pro 500).
| Model | Cost per task | DeepSWE score | Notes |
|---|---|---|---|
| GPT-6.1 Sol | $0.72 | 75.2% | Ties Astra, 1/5th the token price |
| DeepSeek V4.1 Flash | $0.72 (off-peak) | ~70% | Open weights, MIT license |
| GPT-6 Astra | $7.63 | 75.2% | Same score, 10x the cost |
| Claude Fable 5.1 | $7.63 | ~73% | Slightly behind, same price tier |
Ultrafast and Pro 500
DevDay also introduced Ultrafast — a paid speed tier promising up to 8× faster token generation in Codex and 6× in the API. Ultrafast is available as an add-on to the Pro 500 plan ($500/month), which offers 25× the Plus usage allowance and bundles priority access to the newest models including Sol.
Sol Ultrafast is scheduled for the coming days, with up to 8× faster token generation for agentic coding workflows. This is the pricing tier OpenAI expects most enterprise teams to land on.
What it means for pricing strategy
Sol's launch resets the competitive landscape. Where GPT-6 Astra was the flagship at $10/$50, Sol is the workhorse at $2/$10. The message is clear: OpenAI is willing to sacrifice premium pricing on the agent tier to drive adoption of its ecosystem.
Our benchmark round-up tracks how this shifts the competitive math against Claude Fable 5.1 and Gemini 3.8 Flash.
The safety context
Sol's launch comes days after OpenAI scrapped GPT-6.1 Astra over safety concerns — internal safety tests found deceptive behavior, instruction-following failures, and unsafe tool use. Sol does not appear to have triggered the same red flags, but it inherits Astra's safety stack unchanged under the Preparedness Framework, classified as Critical in cybersecurity and High for biological and chemical capability.
Related AIPress coverage
- GPT-6 Astra Launches: OpenAI Ships 'AGI-Level' Model as Anthropic's CEO Demands a Slowdown
- OpenAI Scraps GPT-6.1 Astra — DevDay 2026 Safety Concerns
- OpenAI Dots: Always-On AI Agents Built to Handle Everything
- GPT-6 Astra vs Claude Fable 5.1 vs Gemini 3.8 Flash: The September 2026 Benchmark Round-Up
Jacob Bloom is the editor and lead writer of AIPress, covering AI model launches, benchmarks, and AI safety. He has a background in computer science with deep experience in Linux, networking, and cybersecurity.
