跪拜 Guibai
← All articles
OpenAI · AIGC

GPT-5.6 Lands With a 13-Point Lead Over Claude Fable 5 and a 16x Cost Advantage

By AI袋鼠帝 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The cost-to-performance ratio resets the frontier-model market. A developer who burned through half a $200 Cursor subscription in two days on Fable 5 can now get stronger results on the cheaper Terra or Luna tiers without rationing usage.

Summary

GPT-5.6 arrives as a three-model family: Sol (flagship), Terra (balanced), and Luna (lightweight). On the Agents' Last Exam benchmark, which tests long-chain tasks across 55 professional domains, Sol scores 53.6 against Fable 5's 40.5 — a 13-point gap that dwarfs previous incremental leads. Sol also tops the Artificial Analysis Coding Agent Index, Terminal-Bench, and DeepSWE, using half the tokens and half the time of Fable 5 while costing a third less.

Early integrators report concrete gains. Lovable saw a 25% drop in steps needed to build an app, a 35–48% reduction in tool calls, and 15% fewer stuck sessions. Qodo found token usage fell 3× compared to GPT-5.5 and response latency halved. Design output, a long-standing weakness, now produces interactive games and visualizations from single prompts, and the model uses computer-use capabilities to inspect rendered output and self-correct visual bugs.

Pricing reshapes the competitive landscape. Terra and Luna beat Fable 5 on select evals at roughly 1/16th the cost. Sol itself is priced at $5/$30 per million input/output tokens, with Terra at $2.50/$15 and Luna at $1/$6. All three models are available today via API, ChatGPT Plus/Pro, and Codex, with an ultra mode for Pro and Enterprise users that runs four parallel agents on complex tasks.

Takeaways
GPT-5.6 Sol scored 53.6 on Agents' Last Exam versus Fable 5's 40.5, a 13-point lead.
Sol's medium reasoning mode already beats Fable 5 at roughly one-quarter the cost.
On the Coding Agent Index, Sol hit 80 to Fable 5's 77.2 while using half the tokens, half the time, and costing a third less.
Terra and Luna beat Fable 5 on some evaluations at approximately 1/16th the price.
Lovable reported 25% fewer build steps, 35–48% fewer tool calls, and 15% fewer stuck sessions after integrating GPT-5.6.
Qodo measured 3× lower token consumption than GPT-5.5 and halved response latency for code review.
Frontend and design output improved markedly: single-prompt interactive games, museum sites, and visualizations now render correctly, and the model self-corrects visual bugs via computer use.
ChatGPT Work renders interactive visualizations directly in the dialog box rather than a side preview.
PPT generation remains ugly despite accurate data extraction.
Pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million input/output tokens.
All three models are available today via API, ChatGPT Plus/Pro, and Codex; an ultra mode runs four parallel agents for Pro and Enterprise users.
Conclusions

OpenAI's lead this time is not incremental — a 13-point gap on a long-chain agent benchmark and a 16× cost advantage on mid-tier models together shift the default choice for cost-sensitive agent workloads.

The self-correcting design loop — generate code, render it, inspect the output with computer use, and fix visual bugs — turns a one-shot code generator into something closer to a junior frontend developer who checks their own work.

Rendering interactive visualizations inside the dialog box rather than a side panel is a small UX change that signals where agent interfaces are heading: the conversation itself becomes the workspace.

Anthropic's account-banning practices get a brief but pointed mention, hinting that access friction is becoming a competitive liability when the performance and pricing gap widens this much.

Concepts & terms
Agents' Last Exam
A benchmark that evaluates AI models on long-chain, multi-step tasks across 55 professional domains, measuring agentic reasoning rather than single-turn accuracy.
Terminal-Bench / DeepSWE
Benchmarks that test an AI model's ability to navigate real codebases, execute shell commands, and complete multi-file engineering tasks — closer to real-world software work than isolated coding challenges.
Ultra mode
A GPT-5.6 feature for Pro and Enterprise users that runs four agents in parallel to handle particularly complex tasks, trading concurrency for speed and thoroughness.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗