In three lines
- cost efficiency and general-purpose agent work, GPT‑5.6 Sol has the advantage. Its API output price is 40% lower than Fable 5's.
- large codebases and long-running autonomous coding, Claude Fable 5 is a strong choice. Fable 5 also leads in the published SWE‑Bench Pro results.
- If you handle sensitive data, review the policy differences. Fable 5 requires 30-day data retention for safety monitoring.
Key specifications
The prices below are per one million tokens for each company's own API. Subscription fees and cloud marketplace prices are separate, and different billing terms may apply to caching and long inputs.
| Item | GPT‑5.6 Sol | Claude Fable 5 |
|---|---|---|
| Release | Generally available July 9, 2026 | Released June 9, 2026; access restored July 1 |
| API model name | gpt-5.6-sol |
claude-fable-5 |
| Context window | 1,050,000 tokens | 1,000,000 tokens |
| Maximum output | 128,000 tokens | 128,000 tokens |
| API input price | $5 / 1 million tokens | $10 / 1 million tokens |
| API output price | $30 / 1 million tokens | $50 / 1 million tokens |
| Main access channels | ChatGPT, Codex, OpenAI API | Claude, Claude Code, Claude API, and major cloud platforms |
| Data-related restrictions | Check the retention policy for each product and contract | Fable usage data must be retained for 30 days; ZDR is not supported |
Who leads in published benchmarks?
No single model wins every evaluation. The figures below come from OpenAI's GPT‑5.6 launch materials, where both models can be compared in the same table. Results can vary by evaluation environment and reasoning settings and do not guarantee identical performance in real work.
| Evaluation | GPT‑5.6 Sol | Claude Fable 5 | Leader in this table |
|---|---|---|---|
| Agents' Last Exam | 52.7% | 40.5% | GPT‑5.6 Sol |
| AA Coding Agent Index v1.1 | 80.0 | 77.2 | GPT‑5.6 Sol |
| SWE‑Bench Pro | 64.6% | 80.0% | Claude Fable 5 |
| Terminal‑Bench 2.1 | 88.8% | 83.1% | GPT‑5.6 Sol |
| Toolathlon | 58.0% | 61.7% | Claude Fable 5 |
Sol was strong in broad agent evaluations and terminal tasks, while Fable 5 led in repository-level software engineering and some tool-use evaluations. It is therefore inaccurate to declare an absolute winner from a single benchmark.
Strengths and limitations of each model
GPT‑5.6 Sol: Cost efficiency and broad capabilities for complex tasks
OpenAI describes Sol as a flagship spanning coding, knowledge work, research, science, computer use, and design. Its input price is half that of Fable 5 and its output price is 40% lower, which can create a significant cost difference for agents that need large contexts and repeated calls.
- Strengths: Lower token prices, strong terminal and agent evaluation results, and broad availability through ChatGPT, Codex, and the API
- Limitations: Long inputs exceeding 272,000 tokens are subject to an additional pricing multiplier, and high-risk biological or cyber requests may be refused or undergo additional review
Claude Fable 5: Long-running autonomous coding and complex professional work
Anthropic describes Fable 5 as a model designed for agent tasks that run for days, large-scale migrations, complex implementations, and document-heavy professional work. Its core positioning is long-running work in which it plans, delegates to sub-agents, and verifies its own results.
- Strengths: SWE‑Bench Pro, long-running coding, complex codebases, interpretation of documents, tables, and charts, and self-verification
- Limitations: High token prices, mandatory 30-day data retention, and some cyber or biological requests may be routed to Opus 4.8 or refused
How to choose by use case
When GPT‑5.6 Sol is the better fit
- Services where API cost and throughput matter
- Combining coding with research, document work, and tool use
- Working across ChatGPT and Codex
- Frequently using terminal-based agent automation
When Claude Fable 5 is the better fit
- Autonomously modifying a large codebase over long periods
- Complex migrations and self-verification are essential
- Analyzing professional documents containing PDFs, tables, and charts
- Quality on a difficult individual task matters more than cost
Bottom line: GPT‑5.6 Sol is the more balanced choice for most general-purpose work and cost efficiency. If long-running autonomous coding and handling complex codebases matter more than budget, Claude Fable 5 is worth testing first. Before making a final choice, the most reliable approach is to evaluate both models on the same sample tasks.
What to keep in mind when comparing
The benchmark results were published by the model developers and can vary with prompts, tools, reasoning effort, and scoring methods. Prices may also differ depending on caching, long inputs, regional inference, and cloud provider. This article is based on public information available as of July 19, 2026, and may change as product policies evolve.
Review basis and related tools
Fact check: 2026-08-07. Model names, prices, context limits, and retention terms were checked separately against each vendor's product and API documentation. Benchmark figures are explicitly limited to the common table published by OpenAI and are not treated as a neutral real-world ranking.