GPT‑5.6 Sol vs Claude Fable 5
Sol is cheaper and stronger on published agentic-coding and long-horizon professional results. Fable leads on repo-level SWE‑bench Pro, aggregate intelligence, difficult math, and parts of knowledge work. This is a real split frontier—not a single-score race.
Last updated July 10, 2026 · Published scores, source caveats, no invented rows
GPT‑5.6 Sol
Claude Fable 5
Token economics: Sol costs materially less
Fable is the more expensive model before performance enters the picture. Sol’s input price is half Fable’s, its output price is 40% lower, and its cache reads are half the price. Fable’s advantage is that Anthropic documents standard pricing across its full 1M-token context window; Sol adds a surcharge when input exceeds 272K tokens.
| API property | GPT‑5.6 Sol | Claude Fable 5 |
|---|---|---|
| Input / 1M tokens | $5.00 | $10.00 |
| Output / 1M tokens | $30.00 | $50.00 |
| Cached input / 1M | $0.50 | $1.00 |
| Cache write / 1M | $6.25 | $12.50 for 5-minute cache; $20 for 1-hour cache |
| Batch input / output | $2.50 / $15.00 | $5.00 / $25.00 |
| Context / max output | 1.05M / 128K tokens | 1M / 128K tokens |
| Long-context policy | Over 272K input: 2× input and 1.5× output for full request | 1M context at standard pricing |
| Model access | ChatGPT, Codex, and OpenAI API | Claude API, Claude Code, Claude, AWS, Google Cloud, Microsoft Foundry |
The published split: coding is not one category
OpenAI’s launch table offers the clearest direct comparison. It shows Sol ahead on the Artificial Analysis Coding Agent Index, DeepSWE, and Terminal‑Bench. Fable is dramatically ahead on SWE‑bench Pro. Those evaluations test different things: terminal-driven task completion and long-horizon engineering versus fixing real repository issues. Treat each score as a task-specific signal, and note that the comparison table is vendor-published.
| Coding & agent evaluation | GPT‑5.6 Sol | Claude Fable 5 | Published edge |
|---|---|---|---|
| Artificial Analysis Coding Agent Index v1.1 | 80.0 | 77.2 | Sol +2.8 |
| SWE‑bench Pro | 64.6% | 80.0% | Fable +15.4 pts |
| DeepSWE v1.1 | 72.7% | 69.7% | Sol +3.0 pts |
| Terminal‑Bench 2.1 | 88.8% | 83.1% | Sol +5.7 pts |
| Artificial Analysis Intelligence Index v4.1 | 58.9 | 59.9 | Fable +1.0 |
Professional work: near tie overall, different strengths
The aggregate intelligence picture is close: Artificial Analysis lists Fable at 60 and Sol at 59 at their respective high-effort settings. But the published sub-benchmarks split. Sol leads the OpenAI table’s long-horizon professional workflow and browsing rows; Fable leads GDPval‑AA and HealthBench Professional by narrow margins.
Professional & computer use
| Evaluation | Sol | Fable |
|---|---|---|
| Agents’ Last Exam | 52.7% | 40.5% |
| GDPval‑AA v2 | 1,747.8 Elo | 1,759.6 Elo |
| Management consulting tasks | 43.2% | 35.5% |
| BrowseComp | 90.4% | 84.3% |
| OSWorld 2.0 | 62.6% | 54.8% |
Academic & health
| Evaluation | Sol | Fable |
|---|---|---|
| HealthBench Professional | 60.5% | 60.9% |
| GPQA Diamond | 94.6% | 92.6% |
| FrontierMath Tier 1–3 | 89.0% | 87.0% |
| FrontierMath Tier 4 | 83.0% | 87.8% |
| AutomationBench | 18.1% | 17.4% |
Tool orchestration and safety behavior
Safeguards: Anthropic describes Fable as Mythos-class with always-on adaptive thinking and safety handling for high-risk requests. OpenAI describes Sol’s layered real-time safeguards and differentiated access controls. Do not assume equal refusal behavior or latency from benchmark results—test your own legitimate workflows.
| Tool / operational dimension | Sol | Fable |
|---|---|---|
| MCP Atlas | Not published | 83.3% Anthropic configuration |
| Toolathlon | 58.0% | 61.7% |
| Context pricing | Surcharge over 272K input tokens | Full 1M context at standard pricing |
| Reasoning style | Configurable, including max; ultra coordinates subagents | Always-on adaptive thinking |
Radar: different peaks, not a single winner
The five axes are transparent task dimensions above, not a weighted “best model” score. Fable’s MCP evidence uses its published result; Sol’s missing MCP Atlas row is shown as missing rather than estimated. Cost efficiency is inverted so cheaper extends further.
● GPT‑5.6 Sol ● Claude Fable 5
- Sol’s larger terminal and cost axes are meaningful for agent workflows with many tool turns.
- Fable’s SWE‑bench Pro result is large enough to matter for repo issue-resolution work.
- The one-point aggregate-intelligence gap is too close to be a safe procurement tie-breaker alone.
What independent measurement adds
Artificial Analysis reports Sol at 59 versus Fable at 60 on its Intelligence Index at high effort. At the same time, it reports Sol leading the Coding Agent Index at 80, ahead of Fable’s 77.2, and estimates Sol’s Intelligence Index task cost at $1.04—about one-third of Fable’s. Its commentary is especially useful because it does not collapse this into a sweep: Fable leads its overall intelligence leaderboard, while Sol leads its coding-agent index.
Decision guide
You build terminal-driven or tool-heavy coding agents, care about the published Coding Agent Index / DeepSWE / Terminal‑Bench leads, want lower token economics, or need configurable max and multi-agent-style ultra workflows.
Your priority is real repository issue resolution, especially if SWE‑bench Pro matches your task style; you need the published MCP Atlas result; or Fable’s 1M standard-priced context and adaptive-thinking behavior fit your stack.
You are selecting a flagship for production. Run held-out repository issues, terminal tasks, long-context documents, and tool chains. Measure successful task cost, retries, tool failures, latency, merge acceptance, and human-review load.
Verdict
Sol is the more economical agentic-engineering flagship. It is materially cheaper, publishes stronger terminal and coding-agent results, and leads long-horizon professional workflow rows in OpenAI’s table. Fable remains the stronger repository-resolution and aggregate-intelligence bet on published evidence. Its SWE‑bench Pro lead is decisive, it edges Sol on the Artificial Analysis Intelligence Index, and it has a published MCP Atlas result.
The honest conclusion is a split frontier. Use Sol where agent execution, tool loops, and cost per run dominate. Use Fable where issue resolution, deep analysis, and standard-priced long context dominate. The best implementation is often a router with a real evaluation set—not a permanent ideological commitment to one frontier model.
Sources & methodology
- OpenAI — GPT‑5.6 launch and evaluation tables: Sol pricing and the shared benchmark table used here.
- OpenAI API docs — GPT‑5.6 Sol: context, output limit, caching, and long-context pricing.
- Anthropic — Claude Fable 5 and Mythos 5 and Anthropic model-selection docs: Fable availability, pricing, context, output, and positioning.
- Anthropic pricing documentation: Fable cache, batch, and long-context policy.
- Claude Fable 5 & Mythos 5 System Card: Fable methodology and published figures.
- Artificial Analysis — GPT‑5.6 benchmarks across intelligence, speed and cost: independent index and estimated task-cost context.
- Benchmark results depend on configuration, scaffolding, effort, tool access, and dataset. Vendor-published comparison rows are labeled as such; charts preserve native values and are not composite rankings.