GPT-5.6 Sol vs Claude Fable 5: Why the Scores Split and How to Route Real Work¶

For / Key Points

For: Engineering leads choosing between Codex and Claude Code models, and teams trying to read frontier-model benchmarks without overfitting to one chart.

Key Points:

General intelligence is nearly tied, but Fable 5 leads in analysis quality while Sol leads in execution efficiency and price
The coding split comes from what each benchmark measures: repository repair and terminal autonomy are different tasks
Task-level routing is more practical than standardizing on one model

On July 9, 2026, OpenAI made GPT-5.6 Sol generally available across ChatGPT, Codex, and the OpenAI API.¹ Comparison with Claude Fable 5 started immediately because both models sit in the current frontier tier.

The public data available through July 13, 2026 does not produce a single winner. The useful question is narrower: which work should go to Sol, and which work should go to Fable 5?

The Short Answer¶

The practical split is Fable 5 for quality and Sol for efficient execution. Fable 5 is stronger in human chat preference and deep analytical work. Sol is stronger in the Codex agent stack, command-line work, output polish, and API economics.

Evaluation Axis	Leading Model	How to Read It
General intelligence	Fable 5, narrowly	Artificial Analysis shows 60 vs 59
Human text preference	Fable 5	Ranked first on LMArena Text
Coding Agent Index	Sol	80.0 in a Codex harness
SWE-Bench Pro	Fable 5	Better fit for repository bug-fix tasks
Terminal-Bench 2.1	Sol	Better fit for autonomous CLI loops
Analytical knowledge work	Fable 5	Higher AA-Briefcase analytical quality
Presentation polish	Sol	Highest AA-Briefcase Presentation Elo
API price	Sol	Lower input and output rates

That table rules out a simplistic "best model" decision. Model choice has to follow the artifact, the harness, and the cost envelope.

Independent Benchmarks¶

The intelligence gap is small; the cost and output-profile gaps are not. Artificial Analysis scores Claude Fable 5 at 60 and GPT-5.6 Sol at 59 on its Intelligence Index.²³⁴ On the Coding Agent Index, Sol leads at 80.0 while Fable 5 scores 77.2.²

Metric	GPT-5.6 Sol	Claude Fable 5
Artificial Analysis Intelligence Index	59	60
Coding Agent Index	80.0	77.2
API input price per 1M tokens	$5	$10
API output price per 1M tokens	$30	$50
Intelligence Index output tokens	70M	87M
Output speed	69.3 token/s	59.2 token/s
Time to first token	157.71s	112.79s
Context window	1M	1M

In this measurement, Sol takes longer to produce the first token, then generates faster and uses fewer output tokens across the evaluation. Fable 5 starts earlier but emits more tokens across the evaluation. Output-token volume does not establish depth or quality by itself; the defensible conclusion is narrower: Sol makes API cost and processing volume easier to budget.

Human preference currently favors Fable 5. On the July 13, 2026 LMArena Text Arena, Fable 5 ranked first at 1508±7, while Sol xhigh ranked tenth at 1484±11.⁵ Sol had 2,992 votes versus 7,959 for Fable 5, so the rank should still be treated as early.

Why Coding Scores Split¶

SWE-Bench Pro and Terminal-Bench are not measuring the same kind of coding. In OpenAI's cross-model table, Fable 5 scores 80.0% on SWE-Bench Pro while Sol scores 64.6%. On Terminal-Bench 2.1, Sol scores 88.8% while Fable 5 scores 83.1%.¹

Benchmark	GPT-5.6 Sol	Claude Fable 5	Lead
SWE-Bench Pro	64.6%	80.0%	Fable 5
DeepSWE v1.1	72.7%	69.7%	Sol
Terminal-Bench 2.1	88.8%	83.1%	Sol
GPQA Diamond	94.6%	92.0%	Sol
FrontierMath Tier 4	83.0%	87.8%	Fable 5
AutomationBench	18.1%	17.4%	Sol, narrowly
Toolathlon	58.0%	61.7%	Fable 5
HealthBench Professional	60.5%	60.9%	Essentially tied

Source caveat

This table comes from OpenAI's launch post. It is useful for comparison, but it is still a first-party cross-model evaluation, not a fully independent result.

SWE-Bench Pro rewards reading an existing repository, understanding the issue, narrowing the affected surface, and writing the smallest correct patch. Repository context is the core difficulty.

Terminal-Bench rewards a different loop: run a command, read the output, revise the hypothesis, and run the next command. Persistence and tool use matter more. The public numbers support a working hypothesis that Fable 5 is better for read-heavy repository repair, while Sol is better for execute-heavy autonomous loops. They do not, on their own, isolate a model-intrinsic capability difference.

The other caveat is the harness. Artificial Analysis evaluates Sol inside OpenAI's Codex harness, not as a bare model.² The comparison therefore includes product-stack differences between Codex and Claude Code.

Knowledge Work¶

For documents and analysis, Fable 5 writes the substance; Sol may finish the artifact better. On AA-Briefcase, Fable 5 leads overall, while Sol has the highest Presentation Elo of any model.² That distinction is the practical part.

The internal split is clear.

Fable 5 has a 56% Rubric Score versus Sol at 42%
Fable 5 scores 1764 in Analytical Quality Elo versus Sol at 1592
Sol has the highest Presentation Elo

For design reviews, research memos, and root-cause analysis, Fable 5 is the safer first model. For slides, spreadsheets, and polished deliverables, Sol's finishing quality matters. A useful workflow is to draft the substance with Fable 5, then use Sol to validate, format, and polish.

Caveats Before Reading the Numbers¶

Some Fable 5 results include Opus 4.8 fallback behavior. Anthropic says that requests related to cybersecurity, biology and chemistry, or distillation may be routed to Claude Opus 4.8 when Fable's classifiers fire.⁶ That is why Artificial Analysis labels the evaluated model as "Claude Fable 5 ... Opus 4.8 Fallback."⁴

Anthropic says more than 95% of Fable sessions involved no fallback in early data.⁶ Still, safety-sensitive tasks need to be tested through the actual routing behavior, not just the headline score. This matters most for cyber, bio, chemistry, and distillation-adjacent workloads.

OpenAI's own table also needs careful reading. The cross-model benchmark rows are useful, but they are first-party numbers. Agents' Last Exam is a good example of why: the same OpenAI page says 53.6 in prose and 52.7 in the table.¹ This article does not use that row as a central argument.

Practical Routing¶

Start with routing, not standardization. For cloud and infrastructure teams, the first split can be this simple.

Use GPT-5.6 Sol for
- Multi-file implementation in Codex
- Repeated shell, Git, and test loops
- Infrastructure setup and incident reproduction
- High-volume workloads where API cost matters
- Presentation- or spreadsheet-heavy deliverables
Use Claude Fable 5 for
- Large repository issue fixes
- Implementation from specs or design docs
- Architecture review and root-cause analysis
- Long-form research and synthesis
- Ambiguous requests that need intent recovery

For existing repositories, the public numbers favor Fable 5. For repeated execution, log inspection, and environment recovery, Sol is likely to win more often. For high-value deliverables, a two-stage workflow is sensible: Fable 5 for substance, Sol for validation and polish.

One factor can outweigh the model choice entirely. Prompt design, tool permissions, tests, retry policy, and context-compression behavior can erase a few benchmark points. The most reliable selection method is still a 20- to 50-task internal eval that records success rate, number of corrections, elapsed time, and total cost.

Summary¶

GPT-5.6 Sol and Claude Fable 5 should not be reduced to one leaderboard decision. Fable 5 is stronger when the work is reading, reasoning, and synthesizing. Sol is stronger when the work is running, testing, and finishing.

That split is the more durable lesson. Teams are no longer just choosing a model; they are designing which stage of a workflow goes to which model and harness. Benchmarks are the starting point, not the final answer.

Claude Fable 5 Launch: What General Availability of a Mythos-Class Model Means — How Fable 5's safeguards and access model work
Same Word, Different Mechanism: Claude Code and Codex Context Compaction Compared — Why harness behavior changes model outcomes
Codex vs Claude Code Subagent Architecture — A closer look at agent-stack differences