Artificial Intelligence

GPT-5.6 Lands as an Agent-First Model, Codex Gets Absorbed, and Claude Resets Quotas

By 甲维斯 · Jul 10, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The model race has split into two tiers: Sol and Fable 5 fight for the top across nearly every benchmark, while every other lab competes on price or niche agent scores. For developers picking a daily driver, the decision now hinges on whether your workload is execution-bound (Sol), bug-hunting (Fable), or budget-constrained (Grok).

Summary

GPT-5.6 arrives as three tiered models — Sol, Terra, Luna — priced to mirror the Opus/Sonnet/Haiku structure. Sol pushes terminal-bench and long-running software engineering scores to new highs, but the bigger shift is architectural: an `ultra` mode spins up four parallel agents, and Programmatic Tool Calling lets the model script its own tool-use logic in JavaScript to cut round-trips. The Codex standalone app disappears into ChatGPT (Codex), and a new Work tab gates access to the 5.6 model family inside the main ChatGPT client.

Claude responded by resetting usage quotas mid-cycle, a tactic OpenAI’s Codex lead openly mocked as a sign of pressure. Meta also entered the fray with Muse Spark 1.1, a model that leads on a handful of agent and computer-use benchmarks but trails on raw coding — and prompted Mark Zuckerberg to tweet for the first time in years.

A cross-model benchmark table shows Sol and Fable 5 trading first-place finishes across twelve categories. The practical takeaway: Sol owns execution-heavy agent tasks, Fable leads on hard bug fixes and math, and Grok 4.5 offers the best cost-performance ratio for high-volume work.

Takeaways

— GPT-5.6 ships in three tiers: Sol ($5/$30 per 1M input/output tokens), Terra ($2.5/$15), and Luna ($1/$6), directly copying Claude’s Opus/Sonnet/Haiku pricing structure.

— Sol scores 88.8% on Terminal-Bench 2.1 and 72.7% on DeepSWE, making it the strongest model for terminal-heavy, long-running software engineering tasks.

— The new `ultra` mode coordinates four agents in parallel for multi-module development, security audits, or simultaneous search/code/test workflows.

— Programmatic Tool Calling lets the model write JavaScript to orchestrate multiple tool calls, loops, and parallel requests, reducing the back-and-forth latency of sequential tool use.

— Context window reaches roughly 1.05M tokens with 128K max output, but accuracy in the 512K–1M range drops to about 73.8%.

— The standalone Codex app is gone; it is now ChatGPT (Codex) with a Work tab that gates access to GPT-5.6 models.

— Claude reset user quotas the same day, a move OpenAI’s Codex lead called a sign of competitive pressure.

— Meta’s Muse Spark 1.1 leads on four agent and computer-use benchmarks against Opus 4.8 and GPT-5.5, but trails on coding and was not tested against Sol or Fable 5.

— Benchmark table across twelve categories shows Sol and Fable 5 trading first place; Grok 4.5 leads on cost-performance for high-volume use.

— Explicit cache breakpoints are now supported, with cache reads priced at 10% of normal input cost and a minimum 30-minute retention window.

Conclusions

OpenAI’s decision to gate GPT-5.6 behind a Work tab inside ChatGPT — while keeping Chat mode on 5.5 — segments its 800M-user base into casual and professional tiers without forcing a migration.

Folding Codex into ChatGPT and adding an Invite Friends feature signals OpenAI is prioritizing consumer distribution over developer tooling identity, betting that broad adoption matters more than a separate power-user brand.

Claude resetting quotas the same day GPT-5.6 shipped is a pricing lever, not a product improvement; it buys goodwill with heavy users at zero engineering cost and signals that Anthropic feels the heat.

Meta’s Muse Spark 1.1 benchmarks omit Sol and Fable 5 entirely, which makes the claimed agent leadership a comparison against last-generation models — a pattern that inflates perceived competitiveness.

The benchmark table reveals a clean specialization: Sol wins on execution and terminal tasks, Fable wins on bug-fixing and math, and no single model dominates every category, so model selection is now a workload-matching exercise, not a leaderboard pick.

Concepts & terms

Programmatic Tool Calling

A capability in GPT-5.6 that lets the model write and execute JavaScript to orchestrate multiple tool calls, including loops, parallel requests, and result filtering, instead of the default sequential call-and-respond pattern. It reduces latency in agent workflows.

Ultra mode

A multi-agent execution mode in GPT-5.6 that spawns four parallel agents to work on different facets of a task — such as searching, coding, testing, and verifying — then merges their outputs. Contrasts with `max` mode, which gives a single model more reasoning time.

Cache breakpoints

Explicit markers a developer can set in a prompt to tell the model where to cache context. Cache reads cost 10% of normal input price, and caches persist for at least 30 minutes, making repeated agent tasks and long system prompts cheaper.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗