# OpenAI Unveils GPT-6.1 Sol: A Mid-Tier Model That Delivers Near-Top-Tier Performance at a Fraction of the Cost
OpenAI has introduced GPT-6.1 Sol, a significant upgrade to the mid-tier model in its GPT-6 family. The release targets a clear pain point for developers and enterprises: achieving high-quality results in agentic coding, computer use, and professional knowledge work without paying top-tier pricing. The model is now available through the OpenAI API under the identifier `gpt-6.1-sol`, as well as integrated into ChatGPT Work and Codex.
## What Sets GPT-6.1 Sol Apart?
The headline improvement is economic. GPT-6.1 Sol delivers near-Astra performance on several demanding benchmarks while costing roughly one-fifth of Astra’s standard input and output rates. Input tokens are priced at $2 per million, output tokens at $10 per million, and cached input at just $0.10 per million tokens.
This pricing structure makes GPT-6.1 Sol particularly compelling for workloads that involve repeated context reuse, such as multi-step agent pipelines, where system prompts, tool definitions, and conversation histories are resent on every iteration.
## The GPT-6 Tier Structure
OpenAI now organizes its GPT-6 lineup into three distinct tiers, each serving a different segment of the market:
– **GPT-6 Astra** — The flagship tier at $10 input, $50 output, and $1 per million cached tokens. Recommended for the most challenging research and scientific tasks.
– **GPT-6.1 Sol** — The new mid-tier at $2 input, $10 output, and $0.10 per million cached tokens. A near-Astra alternative at dramatically lower cost.
– **GPT-6 Luna** — The entry-level tier at $0.10 input, $0.50 output, and $0.01 per million cached tokens. Designed for high-volume, cost-sensitive workloads.
The reduced cached input rate is especially relevant for agent-based applications. Agents typically resend the same system prompt, tool schemas, and conversation history across every step of a task. With GPT-6.1 Sol, cached reads now cost just 5% of the uncached input rate, down from 10% on the previous GPT-6 Sol, cutting agent operating expenses significantly.
## Benchmark Performance
OpenAI has released vendor-reported benchmark results across several domains. Competitor figures were sourced from public reports.
### Agentic Coding
On the DeepSWE v1.1 benchmark, GPT-6.1 Sol matches GPT-6 Astra’s coding performance while costing approximately one-fifth as much. It also surpasses the prior GPT-6 Sol’s best score by 6.4 percentage points, achieving this at a lower reasoning effort level.
### Professional Knowledge Work
On GDP.pdf — a benchmark that evaluates answers to questions posed over complex professional PDFs — GPT-6.1 Sol outperforms Claude Opus 5.5 with fallbacks at less than half the cost per task. On AutomationBench 1.0.6, which spans 47 tools across business workflows, Sol scores 2.2 points above Opus 5.5 at medium effort, at roughly one-third the cost. This represents a 4.8-point improvement over the previous GPT-6 Sol.
### Computer Use
On the OSWorld 2.0 offline benchmark, GPT-6.1 Sol gains 7 points over the prior GPT-6 Sol at maximum effort for less than half the cost. It lands within 2.1 points of Astra while costing approximately one-seventh the per-task price.
### Scientific Reasoning
On Terminal-Bench Science 0.1, GPT-6.1 Sol more than doubles GPT-6 Sol’s score at maximum effort. The average cost per task is $5.47, compared with $23.21 for Claude Opus 5.5 and $23.80 for Astra. Astra still leads at 68.1%, and OpenAI recommends it for the hardest scientific research.
### Factuality
At low reasoning effort, the share of responses containing a factual error dropped from 11.4% to 7.7% — a reduction of roughly 32%. This evaluation used deliberately difficult conversations where users had flagged errors in earlier model versions.
## Technical Specifications for Developers
GPT-6.1 Sol comes with a context window of 1,050,000 tokens and supports up to 128,000 output tokens per request. It accepts text and image inputs and produces text output. The knowledge cutoff is April 30, 2026.
The `reasoning.effort` parameter accepts five levels: low, medium (default), high, xhigh, and max. The none and minimal settings are not supported. Developers should use the Responses API for tool calling; Chat Completions works for requests without tool use.
A notable constraint is the 272,000-token threshold: prompts exceeding this limit cost 2x the input and cache rates and 1.5x the output rate for the entire request. Batch and Flex processing modes are 50% cheaper than standard rates, while Fast mode doubles the standard cost. US and EU data residency options are available, though Fast mode is unavailable when EU residency is selected. Fine-tuning is currently not supported for this model.
OpenAI has also announced plans for a GPT-6.1 Sol Ultrafast option within Codex in the near term, promising up to eight times faster token generation compared to standard speed.
—
## Frequently Asked Questions
**Q: How does GPT-6.1 Sol compare to GPT-6 Astra?**
A: GPT-6.1 Sol achieves near-Astra results on most benchmarks while costing one-fifth of Astra’s input and output rates. Astra remains the recommended choice for the most demanding scientific research tasks where every point of accuracy matters.
**Q: What makes GPT-6.1 Sol attractive for agent-based workflows?**
A: Agentic workflows resend the same system prompts, tool schemas, and conversation history on every step. GPT-6.1 Sol’s cached input rate of $0.10 per million tokens — 5% of the uncached rate — significantly reduces the cost of these repeated reads compared to previous models.
**Q: Is GPT-6.1 Sol available in ChatGPT?**
A: Yes, it is available in ChatGPT Work and in Codex, in addition to being accessible via the OpenAI API.
**Q: Can GPT-6.1 Sol be fine-tuned?**
A: No, fine-tuning is not currently supported for this model.
**Q: What is the maximum context window?**
A: GPT-6.1 Sol supports a context window of 1,050,000 tokens with a maximum output of 128,000 tokens.
**Q: How does the pricing compare to GPT-6 Luna?**
A: GPT-6 Luna is priced at $0.10 per million input tokens and $0.50 per million output tokens, making it far cheaper for straightforward tasks. GPT-6.1 Sol sits in the middle, offering substantially better performance on complex benchmarks at a moderate price increase.
**Q: What reasoning effort levels are supported?**
A: The model supports low, medium (default), high, xhigh, and max effort levels. The none and minimal settings are not available.
**Q: Is there a faster version coming?**
A: OpenAI plans to introduce a GPT-6.1 Sol Ultrafast option in Codex within days, with up to 8x faster token generation than the standard speed.
—
## Conclusion
GPT-6.1 Sol represents a strategic move by OpenAI to bridge the gap between its top-tier Astra model and its more affordable Luna tier. By delivering near-Astra quality on agentic coding, computer use, and professional work at one-fifth the cost, it opens the door for broader adoption of powerful AI agents in production environments. The halved cached input rate is a particularly meaningful improvement for any workflow involving multi-step reasoning or repeated context processing.
For teams currently evaluating whether to upgrade from GPT-6 Sol or migrate from GPT-6 Astra for non-critical tasks, GPT-6.1 Sol offers a compelling middle ground — strong capability at a fraction of the flagship price. As the Ultrafast variant rolls out in Codex, the model’s appeal for real-time and latency-sensitive applications is expected to grow further.
Thank you for reading
