Recent comparisons of distinct AI models have positioned OpenAI’s GPT-5.6 Sol—achieving an impressive 88.8% on a premier coding evaluation—against Anthropic’s Claude Fable 5, which garnered an 80.3% rating in software engineering.
Key Insights:
- GPT-5.6 Sol surpasses Terminal-Bench 2.1 with a score of 88.8%, and its Ultra mode elevates this to 91.9%.
- Claude Fable 5 maintains a substantial edge on SWE-Bench Pro at 80.3%, compared to GPT-5.5’s 58.6%.
- At present, Sol is accessible solely through a limited government-sanctioned preview, while Fable 5 re-entered the global market on July 1.
GPT-5.6 Sol’s Benchmark Assertions
OpenAI introduced the GPT-5.6 series on June 26, marking its first update since the April release of GPT-5.5. This iteration has been segmented into three tiers, with Sol reigning as the flagship model.
The organization reports that Sol has achieved an 88.8% score on Terminal-Bench 2.1, an assessment gauging command-line coding agents’ capacity for planning, iterating, and tool coordination.
Notably, an advanced Ultra mode, which activates synchronized subagents to expedite intricate tasks, boosts this score to 91.9%, the pinnacle recorded on the Terminal-Bench chart.
Reviewers who evaluated the comparative charts placed Fable 5 a few percentage points behind Sol in the same terminal evaluation, although the reported scores fluctuated between 83.4% and 84.3%.
Furthermore, on the ExploitBench security suite, Sol reportedly aligns with Mythos-class performance while utilizing approximately one-third of the output tokens—a vital cost efficiency for prolonged agent operations.
However, independent verification of these figures remains elusive for most outside the preview, a limitation acknowledged by reviewers who nonetheless recognized the raw statistics.
Fable 5’s Coding Dominance and Pricing Structure
Fable 5 continues to dominate the benchmark widely regarded as pivotal for autonomous software tasks, boasting an 80.3% score on SWE-Bench Pro, which evaluates real GitHub issue resolutions, while GPT-5.5 lags at 58.6%. OpenAI has yet to publish a figure for GPT-5.6 in this area.
Analysts who identified such disparities across coding, reasoning, and knowledge assessments express skepticism regarding the capability of a single incremental update to bridge these gaps fully.
On the pricing front, Sol is reportedly set at $5 for every million input tokens and $30 for outputs—substantially lower than Fable 5’s $10 and $50 rates.
Some reviewers suggested that a judicious configuration would direct terminal-oriented agents toward Sol, while recommended repository-level corrections would favor Fable 5.
Access remains a pivotal differentiator since Sol is currently limited to approximately 20 government-approved partners, whereas Fable 5 resumed global access on July 1, offering a temporary usage bonus for paid subscribers until July 7.
June has transformed the landscape of model accessibility for both entities into a fluctuating scenario, underscoring every review.
Washington previously mandated the suspension of Fable 5 and its more potent counterpart, Mythos 5, on June 12, citing significant cybersecurity threats, following an Amazon research discovery of a jailbreak that could generate exploit code.
Commerce Secretary Howard Lutnick confirmed the rollback on June 30 after a two-week evaluation, shortly after Mythos 5 quietly returned to a select group of roughly 100 vetted American organizations.
Source link: Yellow.com.
