August 31 Is AI's Deadline Day: Three Model Sunsets Hit at Once

Model deprecations are normally staggered, announced months apart, easy to triage. August 31, 2026 breaks the pattern: three separate deadlines from three separate labs land within the same twenty-four hours, and together they touch an unusually large slice of the production AI stack — the default workhorse model of one frontier lab, the coding tier of another, and the entire legacy API lineage of one of China’s most popular open-model providers.

If your team runs agents or coding pipelines on Claude Sonnet 5, uses OpenAI’s Codex with a ChatGPT account, or still calls Moonshot’s kimi-k2.5 or moonshot-v1 endpoints, tomorrow is not a drill. Here is what changes, why it happened, and what to do about it.

Deadline one: Claude Sonnet 5 loses its introductory pricing

When Anthropic shipped Claude Sonnet 5 on June 30, it came with a promotional rate: $2 per million input tokens and $10 per million output tokens. The company was explicit that this was temporary — through August 31, 2026 — after which standard pricing of $3/$15 takes effect. That deadline arrives tomorrow.

The headline increase is 50% on both axes. But the effective increase for many workloads is larger, and this is the part teams keep missing: Sonnet 5 introduced a new tokenizer, and Anthropic’s own model documentation states that the same input text produces roughly 30% more tokens than it did on Claude Sonnet 4.6. Code-heavy content sees swings between 10% and 35% depending on the workload. A price per token that rises 50%, multiplied by token counts that rise ~30%, means the same job can cost close to double what it did under the promotional rate — a compounding effect that surfaces only if you recalculated costs on the new tokenizer rather than the old one.

Why the increase? Partly normalization: Anthropic’s introductory rates are designed to drive adoption during the evaluation window, then converge to the model’s standard economics. Sonnet 5 sits just below Opus 4.x in capability while remaining the default model for most Claude traffic, and Anthropic has been transparent that standard pricing reflects the model’s actual serving costs, including its longer context handling and agentic features like Claude Code integrations.

Deadline two: GPT-5.4 and GPT-5.4 mini leave Codex

OpenAI announced weeks ago that on August 31, 2026, GPT-5.4 and GPT-5.4 mini will no longer be available in Codex for users signed in with a ChatGPT account. The models remain available through the API and in Codex sessions authenticated with an API key — this is a product-surface removal, not a full model shutdown.

The replacement path is already in place. OpenAI’s GPT-5.6 family — Sol (flagship), Terra, and Luna — went generally available on July 9, and the official migration guidance maps retiring Codex users to gpt-5.6-terra and gpt-5.6-luna. Terra serves as the direct successor for coding-heavy agent work; Luna is the cheaper, faster tier that OpenAI positioned as matching or exceeding GPT-5.4’s benchmarks at a fraction of the cost. OpenAI even cut prices on Terra and Luna in late July, and added a Fast mode for Sol that delivers up to 2.5x standard processing speed — signals that the company wants the migration to feel like an upgrade, not a shove.

The practical wrinkle: teams that built automation around Codex CLI or the macOS app with ChatGPT sign-in will see those pipelines break unless they either switch model selections to the 5.6 family or move to API-key authentication, which preserves GPT-5.4 access for now.

Deadline three: Moonshot kills kimi-k2.5 and moonshot-v1 entirely

The most abrupt of the three. Following the Kimi K3 launch on July 16, Moonshot closed kimi-k2.5 and the moonshot-v1 series to new users almost immediately. Full platform sunset comes tomorrow, August 31 — and unlike the OpenAI change, this is a full retirement, not a surface restriction. Existing integrations stop working.

The migration target is kimi-k3, the 2.8-trillion-parameter mixture-of-experts model (16 of 896 experts active per token) that Moonshot released full open weights for on July 27. K3’s API pricing sits at $3.00 per million input and $15.00 per million output with a 1M-token context window, and the API remains OpenAI-SDK compatible, which softens the migration for most codebases: swapping a model string is often the whole job. One caveat for planners: platform sunsets are not synchronized across providers — kimi-k2.5 disappears from Moonshot and OpenRouter on August 31, but Azure’s hosted copy holds until 2027, so teams pinned to Azure have runway that direct-API users do not.

Context matters here. Kimi K3 was such a demand shock at launch that Moonshot had to suspend new subscriptions within days because GPUs were saturated. Sunsetting the older lineage is partly capacity reallocation: every H800 serving k2.5 is an H800 not serving K3.

Why one day carries three deadlines

Coincidence plays a role — these are three independent product calendars. But the clustering also reflects how the industry’s release cadence has compressed. Introductory pricing windows are now a standard competitive weapon (Google’s Gemini 3.7 Flash intro rate runs through December 31; OpenAI ran promos on Sol through November 21), which means promo expiries accumulate at quarter boundaries. August 31 is the last day of a fiscal quarter for several labs and the natural line for retiring model generations ahead of fall release cycles.

The deeper signal is that model churn has become an operational cost. When Anthropic can reprice a default model +50% overnight, when OpenAI can pull a frontier coding model from a product surface on six weeks’ notice, and when an entire API lineage can vanish in six weeks, the engineering response stops being “migrate when convenient” and becomes “build for portability from day one.”

The migration checklist

For teams with exposure to any of the three deadlines:

1. Audit model strings today. Grep configs, environment variables, and orchestrator defaults for claude-sonnet-5, gpt-5.4, gpt-5.4-mini, kimi-k2.5, and moonshot-v1. Anything found is either already broken after tomorrow or about to cost more.
2. Recount tokens, don’t extrapolate. Run representative payloads through Sonnet 5’s actual tokenizer and rebuild cost models on real counts. Sonnet 4.6 baselines understate September spend.
3. Re-tune output budgets. If you set max_tokens against the old tokenizer’s counts, September invoices will also reflect budgets that silently inflated.
4. Check your auth method on Codex. API-key sessions keep GPT-5.4; ChatGPT sign-in sessions do not. Decide which you actually need before the default breaks mid-pipeline.
5. Verify fallback routing. If you use a router or gateway (OpenRouter, LiteLLM, or in-house), confirm the fallback chain excludes sunsetting models — a router that fails over to kimi-k2.5 on September 1 fails hard.

The bigger picture

Tomorrow’s triple sunset is a snapshot of an industry that now treats model generations the way cloud providers treat instance types: powerful, replaceable, and disposable on a schedule. The winners of this arrangement are users of the newest tiers — GPT-5.6 Luna’s pricing means capability that was frontier-class in spring now costs a fraction of what GPT-5.4 did. The losers are teams whose architecture assumes any particular model string will exist next quarter.

The labs have made their calendars clear. The question for engineering teams is whether their own architecture has a calendar — a standing process that tracks deprecation notices, rehearses model swaps, and treats every “introductory price” as a number with an expiry date. August 31, 2026 is a good day to start.

---

Sources are listed in the article metadata. Pricing and dates verified against Anthropic’s Sonnet 5 documentation, OpenAI’s developer announcements, and Moonshot’s Kimi platform model list as of August 30, 2026.