跪拜 Guibai
← All articles
Artificial Intelligence · AI Programming

GPT-5.6 Lands Tonight as a Three-Tier Model Family, with Sol Ultra Hitting Codex

By 卡卡罗特AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A three-tier model family with a multi-agent mode landing in Codex changes the economics of AI-assisted development: Sol handles the hard problems, Terra runs the daily volume at half GPT-5.5's price, and Luna executes simple tasks at one-fifth Sol's cost. The timing — the exact day Fable 5 exits free access — turns a model launch into a direct subscription play for developers who are already shopping for alternatives.

Summary

GPT-5.6 is not a single model but a tiered product family: Sol for the hardest coding and agent tasks, Terra as a half-price GPT-5.5 replacement for volume workloads, and Luna as a fast, cheap execution layer. Sol's standout feature is an `ultra` multi-agent mode that dispatches parallel sub-agents to analyze code structure, security, and performance before a main agent synthesizes the results — a shift from building stronger single models toward orchestrating virtual teams. The `max` reasoning intensity setting lets users trade latency and cost for deeper analysis on large codebases or complex bugs.

The release timing is pointed. Anthropic's Fable 5, which dominated AI discourse for two weeks, moves from free inclusion in subscriptions to pay-per-use billing on July 7. GPT-5.6's full launch that same day, with comparable capability and no extra charge, targets users who are re-evaluating their subscriptions. For Chinese developers, the contrast is sharper: Anthropic's aggressive bans — escalating from IP blocks to timezone checks — locked many out entirely, while GPT-5.6 arrives with no such restrictions.

Codex gets the strongest Sol Ultra variant for the first time. Previously, ChatGPT Pro's higher-compute GPT-5.5 model never reached Codex; the version developers used there was a weaker tier. Product lead Tibo confirmed Sol Ultra's Codex integration directly, and the community suspects recent GPT-5.5 degradation in Codex was compute being diverted to prepare for this launch.

Takeaways
GPT-5.6 is a three-model tier: Sol (flagship coding/agent), Terra (GPT-5.5 replacement at half price), and Luna (fast execution at one-fifth Sol's output cost).
Sol's `ultra` mode dispatches multiple sub-agents in parallel — one for code structure, one for security, one for performance — then synthesizes findings through a main agent.
A `max` reasoning intensity setting lets users spend more compute for deeper analysis on large codebases, difficult bugs, or complex architecture.
Sol Ultra arrives in Codex for the first time; the higher-compute GPT-5.5 Pro variant was never available in Codex before.
The launch date — July 7 — is the first day Anthropic's Fable 5 moves from free subscription inclusion to paid add-on billing.
Anthropic's Fable 5 ban campaign against Chinese users escalated from IP blocks to timezone detection embedded in Claude Code, locking out many who had just paid for Pro subscriptions.
Frontend UI generation quality, a long-standing weakness of GPT, is reportedly improved in 5.6.
Community speculation holds that recent GPT-5.5 degradation in Codex was compute being reallocated ahead of the 5.6 launch.
Conclusions

Shipping a multi-agent architecture inside a single model product is a paradigm shift: OpenAI is no longer just scaling model size but orchestrating specialized sub-models as a virtual engineering team.

The three-tier pricing structure — Sol at $30/M output tokens, Terra at $15, Luna at $6 — creates a clear upgrade ladder that encourages developers to route tasks by complexity rather than defaulting to the strongest model.

Timing the launch to the exact day a competitor's free access ends is a retention play aimed at the churn window, not a coincidence. The model's capability matters, but the pricing and access terms are the real weapon.

Anthropic's aggressive geo-blocking of Chinese users creates an asymmetric opening: a capable, unrestricted model arriving the same week Fable 5 becomes paid and inaccessible could capture an entire regional developer base in one move.

The admission that Codex never ran the strongest GPT-5.5 variant reveals a product gap that persisted for months — many Pro subscribers were unknowingly using a weaker model in their primary coding environment.

Concepts & terms
Multi-agent mode (ultra)
Sol's ultra mode dispatches multiple specialized sub-agents to analyze different dimensions of a problem in parallel — code structure, security vulnerabilities, performance bottlenecks — then a main agent synthesizes their findings into a single output. This shifts the model from a single reasoning engine to an orchestrator of a virtual team.
Max reasoning intensity
A user-selectable setting that allocates additional compute resources to a query, trading higher latency and cost for deeper reasoning. Useful for large codebase analysis, difficult debugging, or complex architectural decisions where surface-level answers are insufficient.
Three-tier model family (Sol / Terra / Luna)
GPT-5.6 ships as three distinct models at different price points: Sol for the hardest agent and coding tasks ($30/M output tokens), Terra as a cost-effective GPT-5.5 replacement ($15/M output), and Luna as a fast, cheap execution layer for classification and extraction ($6/M output).
Source: juejin.cn ↗ Google Translate ↗ Backup ↗