OpenAI’s GPT-6.1 Sol Pressures High-End Model Pricing

GPT-6.1 Sol is OpenAI’s efficiency-first upgrade to GPT-6 Sol, a model that the company says nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard token prices.

OpenAI has released GPT-6.1 Sol, an efficiency-focused model that the company says nearly matches its flagship GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth of Astra’s standard token prices. The model is available immediately through ChatGPT Work, Codex, and the API, and it introduces a cached-input price of $0.10 per million tokens, a 95% discount from standard input pricing. The launch lands as enterprise buyers increasingly tie model selection to cost-per-task rather than raw benchmark leadership.

The launch continues OpenAI’s practice of segmenting the GPT-6 family by price-performance rather than by capability alone. GPT-6 Astra remains the top-tier model, GPT-6 Sol was the previous efficiency option, and GPT-6.1 Sol now replaces that option with a model that narrows the gap to Astra on several benchmarks while undercutting both predecessors on cost. OpenAI describes GPT-6.1 as the latest model family in the GPT-6 series, with capabilities comparable to Astra and what the company calls an unmatched combination of speed and affordability.

What is GPT-6.1 Sol and how does it fit into OpenAI’s model lineup?

GPT-6.1 Sol is an upgrade to GPT-6 Sol, the model OpenAI positioned as the cost-efficient tier of the GPT-6 generation. According to OpenAI’s announcement, the new model nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices. The system card addendum for GPT-6.1 Sol states that the model delivers capabilities comparable to those of GPT-6 Astra with an unmatched combination of speed and affordability.

The model naming reflects a broader pattern in OpenAI’s lineup: Astra is the flagship, while Sol variants carry the efficiency role. GPT-6.1 Sol is not a new frontier in raw capability, but it is a meaningful step in the cost-performance frontier. The release signals that OpenAI is competing on unit economics, not only on benchmark scores, as agentic applications multiply the number of tokens consumed per completed task.

How does GPT-6.1 Sol compare with GPT-6 Astra on benchmarks?

OpenAI published benchmark results that position GPT-6.1 Sol close to Astra on agentic tasks while widening the gap over GPT-6 Sol. On DeepSWE v1.1, which measures software-engineering agents, GPT-6.1 Sol matches GPT-6 Astra’s score at roughly one-fifth the cost and exceeds GPT-6 Sol’s best score by 6.4 percentage points. On OSWorld 2.0’s offline set, GPT-6.1 Sol comes within 2.1 percentage points of Astra at roughly one-seventh the cost per task, and it outperforms GPT-6 Sol by seven percentage points at less than half the cost.

The science benchmark shows the largest relative gain. On Terminal-Bench Science 0.1, GPT-6.1 Sol more than doubles GPT-6 Sol’s score at maximum reasoning effort while costing less than half as much per task. OpenAI also reported a reduction in factual errors on difficult prompts: at low reasoning effort, the error rate falls from 11.4% for GPT-6 Sol to 7.7% for GPT-6.1 Sol, a reduction of approximately 32%.

Benchmark	GPT-6.1 Sol result	Comparison to GPT-6 Astra	Comparison to GPT-6 Sol	Cost note
DeepSWE v1.1	Matches Astra's score	Match	+6.4 points vs Sol's best	Roughly 1/5 Astra's cost
OSWorld 2.0 offline	Within 2.1 points of Astra	-2.1 points	+7 points vs Sol	Roughly 1/7 Astra's cost per task
Terminal-Bench Science 0.1	More than doubles Sol's score	Not disclosed	>2x Sol at max reasoning	<1/2 Sol's cost per task
Factual error rate (low reasoning)	7.7%	Not disclosed	11.4% to 7.7% (~32% reduction)	N/A

The benchmark table shows a consistent pattern: GPT-6.1 Sol closes most of the gap to Astra on tasks that require an agent to plan, write code, or operate a computer, while delivering cost reductions that compound over long task runs. The DeepSWE result is particularly notable because software engineering agents typically consume large numbers of input and output tokens as they iterate on tests and diffs.

OpenAI’s benchmark methodology distinguishes offline and online evaluation sets for computer use. The OSWorld 2.0 result cited in the announcement is the offline set, which isolates agent performance without live-environment variability. That distinction matters because offline sets are more reproducible for procurement comparisons, while online sets capture real-world execution risks.

What technical choices drive the cost and error-rate improvements?

OpenAI did not disclose the architecture behind GPT-6.1 Sol in the announcement or the system card addendum, so the efficiency gains are best inferred from the public benchmark and pricing data. The sharp improvement at low reasoning effort, where the factual error rate dropped by roughly a third, suggests the model was trained or tuned to make better use of limited inference compute. That type of gain typically comes from data quality, post-training, or inference-time routing rather than from a larger parameter count.

Reasoning effort is a key variable in the GPT-6 series. OpenAI reports results at both low and maximum reasoning effort, and the factual error rate reduction is measured at low reasoning effort on difficult prompts. That suggests GPT-6.1 Sol is more reliable when inference compute is constrained, which is the regime most developers will use for cost-sensitive applications.

The cached-input price of $0.10 per million tokens is a 95% discount from standard input pricing and is 50% less than GPT-6 Sol’s cached input price, according to OpenAI’s API documentation. For agentic workloads, caching matters because agents frequently re-read the same system prompts, tool schemas, and conversation histories across multiple steps. A long agentic task can generate thousands of cached tokens per step, so the discount directly reduces the cost of loop-based applications.

The model also appears designed for professional workloads that mix structured and unstructured inputs. OpenAI’s API documentation describes GPT-6.1 Sol as delivering near-Astra performance at a lower cost for complex coding, computer use, and professional work. The phrase 'near-Astra' is a deliberate positioning choice: it acknowledges the capability gap on some tasks while arguing that the gap is small relative to the price difference.

What does the new pricing mean for developers and enterprises?

The pricing structure changes the unit economics of AI agents. At $2 per million input tokens and $10 per million output tokens, GPT-6.1 Sol is priced well below the flagship tier, and the cached-input discount lowers the effective cost of long-context sessions. For a coding agent that sends the same repository context to the model across dozens of steps, the cached-input price can become the dominant factor in total cost, making the 95% discount more consequential than the headline input price.

Enterprises evaluating models for agentic deployments have historically faced a trade-off between capability and cost. GPT-6.1 Sol narrows that trade-off by matching Astra on the DeepSWE coding benchmark at one-fifth the cost and coming within 2.1 percentage points on OSWorld 2.0 at one-seventh the cost per task. For buyers, the practical question shifts from 'can this model do the task' to 'how much does a successful task run cost,' and the new benchmark data gives them a direct answer.

The release also pressures competitors that sell high-priced flagship APIs. A model that nearly matches a flagship on agentic benchmarks while charging commodity-style prices for cached input creates a new reference point for procurement. Downstream vendors that build coding agents or computer-use tools on top of OpenAI APIs can now offer comparable functionality at lower marginal cost, which may squeeze margins in the application layer even as it expands the addressable market.

The cost-per-task framing also affects how vendors price their own products. An agent builder that pays $2 per million input tokens can run far more task attempts for a fixed budget than one paying flagship prices, which changes experimentation strategies. Teams that previously reserved flagship models for high-value tasks can now route a larger share of traffic to GPT-6.1 Sol without sacrificing much accuracy on coding and computer-use benchmarks.

How is OpenAI positioning the release?

OpenAI frames GPT-6.1 Sol as an efficiency milestone rather than a capability jump. The company’s announcement states that the upgrade nearly matches GPT-6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices. The system card addendum repeats the affordability theme, describing the model as delivering capabilities comparable to those of the most powerful model in the GPT-6 series with an unmatched combination of speed and affordability.

We’re introducing GPT‑6.1 Sol, an upgrade to GPT‑6 Sol that nearly matches GPT‑6 Astra’s intelligence on agentic coding, computer use, and professional work at one-fifth of Astra’s standard input and output token prices.OpenAI

The choice to publish cost-per-task comparisons alongside benchmark scores is itself a signal. OpenAI’s announcement includes cost multiples for DeepSWE, OSWorld 2.0, and Terminal-Bench Science 0.1, framing performance in terms of what a task costs to run, not just how often it succeeds. That framing aligns with the way enterprises budget for agentic workloads, where a 2-point benchmark gap is often less important than a 5x or 7x cost difference.

What are the deployment options for GPT-6.1 Sol?

GPT-6.1 Sol is available today through three surfaces: ChatGPT Work, Codex, and the OpenAI API. The API model identifier is gpt-6.1-sol, and the pricing is $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens, per the API documentation. OpenAI’s deployment-safety addendum indicates the model has gone through the same evaluation process as other GPT-6 models, which is relevant for enterprises with security review requirements.

1. Choose a deployment surface: ChatGPT Work for professional use, Codex for coding agents, or the API for custom applications.
2. Set the model identifier to gpt-6.1-sol in API calls to access the new tier.
3. Design prompts and agent loops to reuse context so the 95% cached-input discount applies.
4. Compare task-level cost using OpenAI’s benchmark data for DeepSWE v1.1, OSWorld 2.0, and Terminal-Bench Science 0.1.
5. Monitor factual error rates on difficult prompts, which OpenAI reports at 7.7% for low-reasoning-effort runs.

The deployment options reflect OpenAI’s strategy of making the model available in both consumer-facing products and developer-facing APIs. The ChatGPT Work integration gives professional users access without requiring API engineering, while the Codex integration targets software engineering agents directly. For enterprises, the API route preserves control over caching, evals, and cost monitoring.

Developers should note that the cached-input discount applies only to cached tokens, not to the first read of a context. Agent frameworks that structure prompts to maximize cache reuse will see the largest savings. OpenAI’s API documentation does not specify cache-eviction policies in the release announcement, so teams with very long contexts should measure cache-hit rates in their own workloads.

What are the safety and governance considerations for GPT-6.1 Sol?

The GPT-6.1 Sol system card addendum, published on OpenAI’s deployment-safety site, states that the model delivers capabilities comparable to GPT-6 Astra. That designation triggers the same safety-evaluation and monitoring framework used for the broader GPT-6 family. Enterprises subject to internal AI governance policies will likely treat GPT-6.1 Sol as a high-capability model even though its price is lower.

What comes next for the GPT-6 family?

GPT-6.1 Sol’s release suggests that OpenAI is iterating on efficiency-focused variants between flagship launches. The GPT-6 generation now spans three tiers: GPT-6 Astra at the top, GPT-6.1 Sol in the middle, and GPT-6 Sol as the entry point. The gap between GPT-6.1 Sol and Astra is narrow enough on several benchmarks that the next efficiency release could plausibly erase it entirely on some tasks, at which point the flagship tier would need a new capability advantage to justify its price.

The 95% cache discount also sets a benchmark for the broader API market. Competitors offering cached input at higher prices will face pressure to match, particularly for agentic workloads where caching is the dominant cost lever. OpenAI’s API documentation lists the discount explicitly, making it easy for developers to calculate savings and for rivals to be measured against it.

For now, GPT-6.1 Sol is the model to beat on cost-adjusted agentic performance. The release provides a concrete data point: near-flagship capability at one-fifth the token price, with benchmark scores that quantify the remaining gap to Astra. The next stage of the frontier-model race may be less about who leads a single benchmark and more about who delivers the lowest cost per successful task.