OpenAI's GPT-5.6 Sol scores 59 on the Artificial Analysis Intelligence Index, one point behind Anthropic's Claude Fable 5 at one-third the cost. The July 10 release introduces three tiered models — Luna, Terra, and Sol — alongside a multi-agent API and programmatic tool calling.
Key facts
- GPT-5.6 Sol scores 59 on AI Index, one point behind Fable 5
- Sol costs $1.04 per task vs. Claude Fable 5 at ~$3.12
- Sol beats Fable 5 by 13.1 points on Agents' Last Exam
- All models have 1M token context, 128K max output
- OpenAI estimates ~30% of SWE-Bench Pro tasks are broken
OpenAI launched GPT-5.6 on July 10, 2026, with three models: Luna ($1/$6 per million input/output tokens), Terra ($2.50/$15), and Sol ($5/$30). According to Simon Willison, all three have a February 16, 2026 knowledge cutoff, a million-token context window, and 128,000 maximum output tokens.
Benchmarks vs. Claude: price-performance inversion
The flagship Sol scores 59 on the Artificial Analysis Intelligence Index, one point behind Claude Fable 5. Per The Decoder, at $1.04 per task it costs a third of what Anthropic's top model charges. On Agents' Last Exam, a benchmark of long-running professional workflows across 55 fields, Sol sets a new high of 53.6, eclipsing Fable 5 by 13.1 points. Even at medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost.
OpenAI also preemptively attacked a benchmark where it trails badly: SWE-Bench Pro, where Fable 5 scores 80% versus Sol's 64.6%. An OpenAI blog post published the day before estimated that ~30% of SWE-Bench Pro tasks are broken and advised careful examination of results.
API features: multi-agent and programmatic tool calling
The model guidance reveals several new API capabilities. Programmatic Tool Calling allows the model to compose and run JavaScript that orchestrates tool calls, reminiscent of Anthropic's dynamic filtering mechanism. Multi-agent lets the model spin up subagents for parallel focused work — the sub-agent pattern now baked into the core API. Prompt cache breakpoints bring explicit cache control similar to Claude's approach.
ChatGPT Work and Chinese user response
OpenAI also launched ChatGPT Work, an agent-based product powered by Codex and GPT-5.6 that handles complex projects across Google Drive, Slack, and Salesforce. According to SCMP, Chinese users accessing the blocked service via VPNs praised the efficiency, though GPT-5.6 remains more expensive than local rivals: Zhipu AI's GLM-5.2 charges ~$1.40 per million input tokens, and DeepSeek V4 up to $0.44.
What to watch
Watch for third-party agentic coding benchmarks (SWE-bench Verified, Terminal-Bench 2.1) over the next two weeks to validate whether GPT-5.6 Sol's Agents' Last Exam lead holds in head-to-head coding tasks. Also track ChatGPT Work adoption metrics and enterprise seat counts.
Source: scmp.com
[Updated 10 Jul via the_decoder]
Simon Willison noted in early access that GPT-5.6 Sol, while competent, hasn't yet matched Claude Fable 5 on complex coding tasks [per Simon Willison]. He also highlighted that all three models share a February 16, 2026 knowledge cutoff, and that OpenAI's livestream included a demo of 3D pelicans riding various animals, underscoring the model's creative capabilities.
