GPT-6.1 Sol: OpenAI's Near-Astra Model at One-Fifth the Cost for Agentic Coding, Computer Use, and Professional Work

Executive Summary

On September 29, 2026, OpenAI announced GPT-6.1 Sol at DevDay in San Francisco — just one week after the initial GPT-6 Sol and Luna release on September 22. GPT-6.1 Sol is an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work while costing one-fifth of Astra's standard input and output token prices.

At $2 per million input tokens and $10 per million output tokens — the same price as GPT-6 Sol — GPT-6.1 Sol delivers dramatic improvements across its benchmark suite: 75.2% on DeepSWE v1.1 (matching GPT-6 Astra at high effort), 71.4% on OSWorld 2.0 offline (within 2.1 percentage points of Astra at max effort), 31.7% on AutomationBench (2.2 points above Claude Opus 5.5 at medium effort), and more than double GPT-6 Sol's score on Terminal-Bench Science 0.1 at less than half the cost per task. The model supports a 1.05M-token context window, configurable reasoning effort (low through max), Ultrafast mode for up to 8× faster token generation, and multi-agent delegation via the Responses API.

What makes GPT-6.1 Sol notable is not just the raw score improvements — it's the strategic compression of the capability-cost curve. By delivering near-Astra performance at 1/5th the cost, OpenAI has made flagship-level agentic work economically viable for production workloads that were previously restricted to the most expensive tier. Combined with the $0.10 cached input rate (95% discount), this creates a new sweet spot for long-running agents that reuse context across many requests.

---

1. What Is GPT-6.1 Sol?

1.1 The Mid-Tier Frontier Model

GPT-6.1 Sol sits in the middle of OpenAI's GPT-6 family, between the flagship GPT-6 Astra and the cost-efficient GPT-6 Luna:

Model	Positioning	Input Price	Output Price	Cached Input	Context Window
GPT-6 Astra	Flagship — most demanding work	$10/M	$50/M	$1/M	1.05M
GPT-6.1 Sol	Near-Astra — complex work at lower cost	$2/M	$10/M	$0.10/M	1.05M
GPT-6 Sol	Complex coding & agentic workflows	$2/M	$10/M	$0.20/M	1.05M
GPT-6 Luna	Focused tasks at scale	$0.10/M	$0.50/M	$0.01/M	1.05M

The key differentiator between GPT-6.1 Sol and GPT-6 Sol is not pricing — they cost the same — but capability. GPT-6.1 Sol was trained to close the gap with Astra on the tasks that matter most for production: coding, computer use, document understanding, and multi-step business workflows.

1.2 Training Methodology

GPT-6.1 Sol uses the same types of data and training as GPT-6 Astra, as documented in the GPT-6 Astra system card. Unlike the incremental updates seen in previous generations, GPT-6.1 Sol represents a substantial capability upgrade within the same price tier:

"GPT-6.1 Sol delivers substantial improvements over GPT-6 Sol across complex professional tasks, from writing and debugging code to understanding documents and executing multi-step business workflows."

The model supports the full reasoning effort spectrum from low through max, with medium as the default. Notably, unlike GPT-6 Sol, GPT-6.1 Sol does not support none or minimal reasoning efforts — it is designed for work that requires genuine reasoning, not simple text completion.

1.3 Model Properties

Property	Value
Model ID	gpt-6.1-sol
Context window	1,050,000 tokens
Max input tokens	922,000
Max output tokens	128,000
Knowledge cutoff	April 30, 2026
Input modalities	Text, Image
Output modalities	Text
Reasoning efforts	low, medium (default), high, xhigh, max
Unsupported efforts	none, minimal
Fine-tuning	Not supported
Predicted outputs	Not supported

---

2. Benchmark Performance

2.1 Coding: DeepSWE v1.1

On DeepSWE v1.1, which evaluates complex software-engineering tasks in real codebases, GPT-6.1 Sol achieves its most striking result:

Model	Effort	DeepSWE v1.1	Cost per Task
GPT-6.1 Sol	High	75.2%	~$0.27
GPT-6 Astra	High	~75.2%	~$1.35
GPT-6 Sol	Max	68.8%	~$0.27
Claude Opus 5.5	Max	—	~$3.00

GPT-6.1 Sol matches GPT-6 Astra at roughly one-fifth of the cost while eclipsing GPT-6 Sol's best score by 6.4 percentage points at a lower reasoning effort and cost. This is the most significant result in the benchmark suite — it means developers can use GPT-6.1 Sol for production coding agents without sacrificing quality.

2.2 Document Understanding: GDP.pdf

On GDP.pdf, which measures how accurately models answer professional questions using complex PDF documents (including tables, charts, diagrams, and fine-print details):

Model	Effort	GDP.pdf	Cost per Task
GPT-6 Astra	—	State-of-the-art	Highest
GPT-6.1 Sol	—	Near-Astra	~1/5 of Astra
Claude Opus 5.5	With fallbacks	Below GPT-6.1 Sol	~2× GPT-6.1 Sol

GPT-6.1 Sol scores higher than Opus 5.5 with fallbacks at less than half the cost per task across tested reasoning settings, while approaching Astra's state-of-the-art performance at roughly one-fifth the cost.

2.3 Agentic Workflows: AutomationBench

On AutomationBench, which measures whether agents correctly complete multi-step business workflows:

Model	Effort	AutomationBench	Cost per Task
GPT-6.1 Sol	Medium	31.7%	~1/3 of Opus 5.5
Claude Opus 5.5	Medium	29.5%	Baseline
GPT-6 Sol	Medium	26.9%	—

GPT-6.1 Sol scores 2.2 percentage points above Opus 5.5 at medium reasoning effort, at roughly a third of the cost. The score is also up 4.8 percentage points from GPT-6 Sol at the same setting.

2.4 Computer Use: OSWorld 2.0

On OSWorld 2.0's offline set, which evaluates agents on demanding computer-use workflows:

Model	Effort	OSWorld 2.0 Offline	Cost per Task
GPT-6 Astra	Max	73.5%	Highest
GPT-6.1 Sol	Max	71.4%	~1/7 of Astra
GPT-6 Sol	Max	64.4%	—

GPT-6.1 Sol outperforms GPT-6 Sol by seven percentage points at maximum reasoning effort at less than half the cost. It comes within 2.1 percentage points of Astra's score at roughly one-seventh the cost per task.

2.5 Scientific Workflows: Terminal-Bench Science 0.1

On Terminal-Bench Science 0.1, which evaluates scientific workflows including data analysis, simulation, and theorem proving:

Model	Effort	Terminal-Bench Science	Cost per Task
GPT-6 Astra	Max	68.1%	$23.80
GPT-6.1 Sol	Max	~34% (2× GPT-6 Sol)	$5.47
Claude Opus 5.5	Max	—	$23.21
GPT-6 Sol	Max	~17%	~$10.94

At maximum effort, GPT-6.1 Sol costs $5.47 per task on average, compared with $23.21 for Opus 5.5 and $23.80 for Astra — delivering substantial scientific capability at over 75% lower cost than either model. Astra still achieves the highest score and should be used for the most difficult scientific research tasks.

2.6 Factuality

GPT-6.1 Sol improves factual accuracy on difficult prompts:

Model	Effort	Factual Error Rate	Cost per Task
GPT-6.1 Sol	Low	7.7%	Lowest
GPT-6 Sol	Low	11.4%	—
GPT-6 Astra	Low	~5.8%	Highest

At low reasoning effort, GPT-6.1 Sol reduces the share of responses containing a factual error from 11.4% to 7.7% — a reduction of approximately 32%. Across tested reasoning settings, its error rate remains within 1.9 percentage points of GPT-6 Astra's, at less than one-fifth the cost per task.

---

3. Pricing & Cost Economics

3.1 Standard Pricing

Metric	GPT-6.1 Sol	GPT-6 Astra	GPT-6 Luna	Ratio vs. Astra
Input	$2/M	$10/M	$0.10/M	1/5
Cached input	$0.10/M	$1/M	$0.01/M	1/10
Cache writes	$2.5/M	$12.5/M	$0.125/M	1/5
Output	$10/M	$50/M	$0.50/M	1/5

The cached input rate of $0.10/M is the killer feature for agents. At 95% less than standard input pricing and 50% less than GPT-6 Sol's cached input ($0.20/M), this gives developers significant room to build and run capable agents that reuse context across requests.

3.2 Processing Tiers

Tier	Price Multiplier	Use Case
Standard	1×	Default production
Fast	2×	Latency-sensitive apps
Ultrafast	6×	Rapid coding iterations (up to 8× faster)
Batch	0.5×	Off-line processing
Flex	0.5×	Cost-optimized workloads

Ultrafast mode, available for GPT-6.1 Sol, provides up to 8× faster token generation compared to standard speed — a significant advantage for interactive coding workflows in Codex.

3.3 Long-Context Pricing

Prompts with more than 272K input tokens are priced at:

• 2× input and cache rates
• 1.5× output rate

This applies to the full request, not just the tokens above the threshold.

---

4. Tool Support & Capabilities

4.1 Supported Tools (Responses API)

GPT-6.1 Sol supports the full suite of OpenAI tools:

Tool	Description
web_search	Real-time web search
file_search	Search uploaded documents
image_generation	Generate images from text
code_interpreter	Execute code in sandbox
hosted_shell	Access to a hosted shell environment
apply_patch	Apply code patches to files
skills	Custom skill definitions
computer_use	Interact with desktop/browser applications
mcp	Model Context Protocol integration
tool_search	Discover available tools

4.2 Multi-Agent Delegation

GPT-6.1 Sol supports multi-agent workflows in the Responses API (currently in beta). It can assign independent work to subagents and combine their findings into a final response:

4.3 Mid-Turn Steering

Via the Responses WebSocket API, users can send steering messages while the model works. Updates are queued and don't cancel running tools or undo completed actions — a critical feature for long-running agentic tasks.

4.4 Asynchronous Tool Calling

GPT-6.1 Sol supports async tool calling, allowing the model to continue independent work while slower tasks (like running tests) execute in the background. The model waits for results before starting dependent work.

---

5. Safety & Alignment

5.1 Capability Classification

GPT-6.1 Sol is classified as High capability in the Biological and Chemical domain. Its reported results did not cross the indicative Critical thresholds.

5.2 Safety Improvements

GPT-6.1 Sol shows substantial improvements over GPT-6 Sol in alignment evaluations:

• More transparent about limitations — better at recognizing and communicating when it lacks information
• More reliable at respecting user intent — lower failure rates on respecting explicit restrictions
• Better at avoiding unauthorized outcomes during agentic tasks
• No attempts to bypass automated safety reviewers — matching GPT-6 Astra and GPT-6 Sol
• Improved biology refusal — performs comparably to GPT-6 Sol on severe and dual-use prompts while refusing fewer benign prompts

5.3 Cybersecurity

On cybersecurity safety evaluations, GPT-6.1 Sol outperforms all previous models in production-chat evaluations. Compared to GPT-5.6 Sol, it shows modest regressions in synthetic and semi-synthetic agentic environments, but model refusal remains one layer of a defense-in-depth safety stack.

---

6. Integration with the GPT-6 Ecosystem

6.1 ChatGPT Deployment

As of October 7, 2026, GPT-6.1 Sol is available in ChatGPT Work and Codex for all Plus, Pro, Business, Enterprise, and Edu users. It is not yet available in the Chat tab (which uses GPT-6 Sol for paid tiers and GPT-6 Luna for free tiers).

6.2 Intelligent UI

The GPT-6 family introduces Intelligent UI, allowing ChatGPT to compose responses using text, visuals, and interactive elements. While GPT-6.1 Sol is not yet in Chat, the underlying model capabilities support this feature.

6.3 Progressive Answering

GPT-6 models can interleave thinking with answering — beginning to respond while still reasoning. In internal evaluations, GPT-6 Extra High can begin answering in the same amount of time as GPT-5.6 Medium while achieving a better overall score than GPT-5.6 Extra High.

---

7. Comparison with Recent Frontier Releases

7.1 vs. Claude Haiku 5.5

The Claude Haiku 5.5 release (October 7, 2026) represents the opposite end of the spectrum: ultra-low-cost small-model inference. See Claude Haiku 5 5 Cheapest Fastest Small Model Adjustable Effort 1m Context Subagent Browser Use 2026 10 08 for details.

Dimension	GPT-6.1 Sol	Claude Haiku 5.5
Positioning	Near-flagship mid-tier	Cheapest small model
Input price	$2/M	$0.10/M
Output price	$10/M	$0.50/M
Context	1.05M	1M
Best for	Complex coding, computer use	High-volume subagent tasks
Adjustable effort	low–max	low–high (first Haiku with effort)

These models are complementary: GPT-6.1 Sol for the orchestrator/complex tasks, Haiku 5.5 for the subagents/volume tasks.

7.2 vs. Grok 4.7

Grok 4.7 (September 21, 2026) competes in the same mid-tier space with $2/$6 pricing. See Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06 for details.

Dimension	GPT-6.1 Sol	Grok 4.7
Input price	$2/M	$2/M
Output price	$10/M	$6/M
Context	1.05M	500K
DeepSWE	75.2%	71.0%
Computer use	Full support	Grok Bot native
Multi-agent	Responses API (beta)	Grok Bot harness

Grok 4.7 undercuts on output pricing ($6 vs $10) but GPT-6.1 Sol leads on context window and coding benchmarks.

7.3 vs. Qwen3.8-Omni-Flash

Qwen3.8-Omni-Flash (September 18, 2026) targets the omnimodal space at $0.15/$0.47. See Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05 for details.

Dimension	GPT-6.1 Sol	Qwen3.8-Omni-Flash
Positioning	Text + image reasoning	Native omnimodal (text + image + audio + video)
Input price	$2/M	$0.15/M
Output price	$10/M	$0.47/M
Audio/Video	No	Yes (input)
Open weights	No	Yes (Flash-Next architecture)

Qwen3.8-Omni-Flash dominates on multimodal input and price; GPT-6.1 Sol dominates on coding and agentic workflow depth.

---

8. Implementation Guidance

8.1 When to Use GPT-6.1 Sol

Choose GPT-6.1 Sol when:

• You need near-Astra quality for coding, computer use, or professional work
• Cost matters but you can't sacrifice capability
• Your agents reuse context across many requests (cached input at $0.10/M is a game-changer)
• You need multi-agent delegation for complex tasks
• You're running long-running scientific workflows (75% cheaper than Astra/Opus on Terminal-Bench)

Choose GPT-6 Astra instead when:

• You need the absolute highest score on the hardest tasks
• The marginal quality improvement justifies 5× the cost

Choose GPT-6 Luna instead when:

• You need high-volume, focused tasks (classification, extraction, summarization)
• Cost is the primary constraint

8.2 Code Example: Multi-Agent with GPT-6.1 Sol

from openai import OpenAI

client = OpenAI()

# Orchestrator: GPT-6.1 Sol delegates to subagents
response = client.responses.create(
    model="gpt-6.1-sol",
    input=[
        {
            "role": "user",
            "content": "Investigate the authentication bug in our codebase. "
                       "Check the auth module, run tests, and propose a fix."
        }
    ],
    reasoning={"effort": "high"},
    tools=[
        {"type": "code_interpreter"},
        {"type": "file_search"},
        {"type": "computer_use"},
    ],
    # Multi-agent delegation (beta)
    max_agent_steps=10,
)

print(response.output_text)

8.3 Cost Optimization Tips

1. Maximize cache hits: Put stable instructions and reference material before changing task details
2. Use Batch/Flex for offline work: 50% of Standard rates
3. Match effort to task: Start with medium, increase to high only when needed
4. Use Ultrafast for interactive workflows: 6× price but up to 8× faster — net positive for latency-sensitive apps
5. Monitor cache diagnostics: Use the caching dashboard to identify where reuse breaks down

---

9. Key Takeaways

1. Near-Astra at 1/5th the cost: GPT-6.1 Sol matches Astra on DeepSWE v1.1 and comes within 2.1 points on OSWorld 2.0 at dramatically lower cost.
2. Cached input at $0.10/M: The 95% discount on cached input makes long-running agents economically viable.
3. Multi-agent ready: Native support for delegation via the Responses API enables complex workflow orchestration.
4. Ultrafast mode: Up to 8× faster token generation for interactive coding in Codex.
5. Complementary to the small-model tier: Works alongside GPT-6 Luna and Claude Haiku 5.5 in multi-tier agent architectures.
6. Safety improvements: Substantially better alignment than GPT-6 Sol, with no safety reviewer bypass attempts.
7. Not yet in Chat: Available in Work and Codex; Chat rollout pending.

---

10. References & Resources

Official Sources

• Introducing GPT-6.1 Sol — OpenAI announcement
• GPT-6.1 Sol Model Documentation — API docs
• GPT-6.1 Sol System Card Addendum — Safety evaluation details
• A Practical Guide to Building with GPT-6 — Implementation guide
• GPT-6 and Intelligent UI for Everyone — ChatGPT rollout announcement
• OpenAI Mathematics Repository — 719 manuscripts from internal frontier model

Related Research

• Claude Haiku 5 5 Cheapest Fastest Small Model Adjustable Effort 1m Context Subagent Browser Use 2026 10 08 — Claude Haiku 5.5 small-model tier
• Grok 4 7 Frontier Coding Knowledge Work New Safeguard Stack Longer Rl 2026 10 06 — Grok 4.7 frontier coding model
• Qwen3 8 Omni Flash Native Omnimodal Agentic Media 1m Context 2026 10 05 — Qwen3.8-Omni-Flash omnimodal model

---

11. Future Directions

11.1 ChatGPT Integration

GPT-6.1 Sol is expected to roll out to the ChatGPT Chat tab in the coming weeks, replacing GPT-6 Sol for paid tiers. This would bring near-Astra intelligence to the everyday conversation experience.

11.2 Fine-Tuning Support

Fine-tuning is not yet supported for GPT-6.1 Sol. Given the model's strong performance on coding and professional tasks, fine-tuning support would enable domain-specific optimization at mid-tier pricing.

11.3 The Mathematics Frontier

OpenAI's release of 719 mathematical manuscripts from an internal frontier model (October 6, 2026) demonstrates the research capabilities of the GPT-6 family. The model that produced these results — described as the same internal model behind the Navier-Stokes solution — has not yet been publicly released. GPT-6.1 Sol represents the current public frontier, but the gap between internal and public capability remains significant.

11.4 The Mid-Tier Race

With GPT-6.1 Sol, Grok 4.7, and the upcoming Mistral Large 4 all competing in the mid-tier space, the next quarter should see aggressive capability improvements at stable pricing. The winner in this segment will be the model that best balances coding capability, computer use, and cost for production workloads.

---

GPT-6.1 Sol: Near-flagship intelligence at mid-tier pricing. The new sweet spot for production agents.