# GPT-5.6 Sol Beats Fable 5 on DeepSWE at One-Third the Cost

GPT-5.6 Sol scored ~72% on the DeepSWE coding-agent benchmark to Claude Fable 5's ~70%, at one-third the cost. The cheaper model leads the first credible third-party comparison since both reached GA.

By Neomanex. Updated 2026-07-12.

**[GPT-5.6 Sol](/models/gpt-5-6-sol) has beaten [Claude Fable 5](/models/fable-5) on DeepSWE — the most rigorous coding-agent benchmark available — while costing roughly one-third as much per task. The numbers arrive just as Fable 5's subscription-included access window closes tonight.**

DeepSWE is not a code-generation contest. It drops an agent into a real open-source repository with a requested change and verifies whether the final patch actually works — 113 tasks across 91 repositories and 5 languages, testing repo navigation, bug diagnosis, code editing, tool use, and multi-step repair. The score means "what percentage of real software engineering tasks were solved." GPT-5.6 Sol solved more of them, for less money.

## What happened: GPT-5.6 Sol tops DeepSWE benchmark at one-third the cost

Independent evaluator Rohan Paul published the first credible third-party DeepSWE benchmark comparison since both models reached general availability. GPT-5.6 Sol (max) scored approximately 72–73% at roughly $8.40 per task, while Claude Fable 5 (max) scored approximately 70% at $13–22 per task ([Rohan Paul](https://www.rohan-paul.com/p/gpt-56-beats-fable-5-by-on-deepswe), 2026).

| Model | DeepSWE score | Cost per task | Input / Output pricing |
|-------|--------------|---------------|----------------------|
| GPT-5.6 Sol (max) | ~72–73% | ~$8.40 | $5 / $30 per 1M tokens |
| Claude Fable 5 (max) | ~70% | $13–22 | $10 / $50 per 1M tokens |

The Artificial Analysis Coding Agent Index — which combines DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA across model-specific harnesses — corroborates the finding: GPT-5.6 Sol (max) in Codex leads with 80 points, ahead of Claude Fable 5 (max) in Claude Code at 77 ([Artificial Analysis](https://artificialanalysis.ai/articles/gpt-5-6-has-landed), 2026).

On the broader Intelligence Index, Sol scores 59 to Fable 5's 60 — one point behind at **one-third the cost per task** ($1.04 vs $2.75).

The cost gap across Anthropic's lineup is even starker:

- GPT-5.6 Sol: **3x cheaper than Fable 5**
- GPT-5.6 Sol: **5–7x cheaper than Opus 4.8**
- GPT-5.6 Sol: **10x cheaper than Sonnet 5**

These multipliers come from the [International Cyber Digest](https://x.com/IntCyberDigest/status/2075754446152466748) (2026) analysis, which also tracks Terra at 2x and Luna at 2x cheaper than Fable 5.

## Why it matters

GPT-5.6 Sol has been in the directory at a Pending verdict since its June 26 government-gated preview. The independent benchmark data the directory was waiting for has now arrived — and Sol has delivered.

Fable 5's subscription-included access expires tonight at 11:59 PM PT. After that, it's credit-only at $10/$50 — Anthropic's most expensive publicly available model pricing. There is no announced return to subscriptions.

GPT-5.6 Sol launches at $5/$30, subscription-included, and now carries a confirmed coding-agent benchmark lead.

The government-gating drama that surrounded Sol's initial preview obscured a simpler competitive reality: when both models are available, one costs dramatically less and solves more real coding tasks. The policy question was always a sideshow. The benchmark is the main event.

For teams building coding agents, the cost-per-task math compounds fast. A 50-task overnight agent run at $22/task (Fable 5) costs $1,100. The same workload on Sol costs roughly $420. Over a month of daily runs, that gap is **$20,000**.

## What changes for you

If you are choosing between GPT-5.6 Sol and Claude Fable 5 for coding-agent workloads, the directional signal is unambiguous. Sol delivers measurably more solved tasks at one-third the cost. Run your own representative repos to confirm, but do not wait for a tie-breaker that already arrived.

If Fable 5 is in your pipeline, audit **today**. After 11:59 PM PT, every agent loop burns credits at $10/$50 with no metered subscription fallback. Budget accordingly — there is no automatic cap on spend.

For teams that do not need max reasoning on every task: GPT-5.6 Terra (77 on the Coding Agent Index at ~$0.55/task) and Luna (75 at ~$0.21/task) offer even better value for routine coding work.

## FAQ

**Is DeepSWE a reliable benchmark?**
DeepSWE tests agents inside real open-source repositories with program-based verifiers — not multiple-choice questions or function-completion tasks. It measures whether a patch actually works in practice. The benchmark spans 113 tasks, 91 repositories, and 5 languages, making it more representative of real-world engineering than SWE-bench Verified, which relies on unit-test pass/fail.

**Does this mean Sol is the better model for everything?**
No. On AA-Briefcase (knowledge-work tasks), Fable 5 still leads with a Rubric Score of 56% vs Sol's 42%. On the Intelligence Index, the two are one point apart. The data says Sol is the better coding agent for the money — not that Sol is universally superior. For analytical reasoning or presentation-quality outputs, Fable 5 remains competitive.

**What about GPT-5.6 Terra and Luna?**
Both score respectably on the Coding Agent Index — Terra at 77 and Luna at 75 — at even lower per-task cost. If your coding tasks don't require max reasoning effort, the cheaper variants may offer better value still. The entire GPT-5.6 family now defines a new Pareto frontier on intelligence-vs-cost charts.

Canonical: https://neomanex.com/news/gpt56-sol-beats-fable5-deepswe-benchmark