Six model sunsets and price changes this summer: one page, kept current

The short version: between July 20 and August 31, 2026, at least six model shutdowns and price changes take effect, and the nearest one is the day after tomorrow. After July 24, 15:59 UTC, any code still calling deepseek-chat starts failing in bulk. This is the duty roster we keep for fiction-continuation users; every line was checked against the official announcement (verified 2026-07-18). Table first.

Already in effect: Gemini's free tier and the 2.0 shutdown

Two entries are history, and they set the tone. On April 1, Google removed Gemini 2.5 Pro, 3 Pro, and 3.1 Pro from the API free tier, leaving only Flash and Flash-Lite with reduced quotas — a change that shipped with no formal changelog. Developers found out from “This model requires a billing-enabled project” errors. Then on June 1, the official Gemini API changelog recorded the shutdown of four Gemini 2.0 models (gemini-2.0-flash, -flash-001, -flash-lite, -flash-lite-001), pointing users to 3.5 Flash or 3.1 Flash Lite. The pattern to internalize: free tiers tighten, old model names die, and the bell does not always ring in advance.

Mid-July: DeepSeek's peak-hour pricing — a time-of-day rate card for LLM tokens

On June 29, DeepSeek emailed API subscribers: the V4 official release lands mid-July, and with it comes peak/off-peak pricing. During two daily windows — 9:00-12:00 and 14:00-18:00 Beijing time — every billing item charges double the off-peak rate; off-peak stays at the current list price (TechNode’s report; DeepSeek says it will email 24 hours before billing switches). Note: every day, weekends included. As of our July 18 check, the official pricing page still shows a single rate (V4 Flash: $0.14 per million input tokens on cache miss, $0.28 output) with no peak column — treat the email in your inbox as the switch signal.

For fiction users the practical impact is smaller than the headline. If you write in the evening, you never leave the off-peak rate. What needs moving is unattended batch work — bulk generation, whole-book preprocessing — that currently lands inside the peak windows. Shift it a few hours and your bill does not change.

July 20: Fable 5’s free window closes

Claude Fable 5 came back online July 1 with a promotional window for Pro, Max, Team, and select Enterprise plans — up to 50% of weekly usage limits at no extra cost. The end date slipped twice and finally landed on July 19, 11:59:59 PM Pacific. From July 20, Fable 5 is no longer included in any subscription; continued use requires enabling usage credits in Claude’s settings, billed at $10 per million input tokens and $50 per million output, with cache hits at $1. The API was always metered at this same rate card, so API users see no change on this date.

Whether that price is worth paying for fiction is a real question — it is the most expensive generally available model on the market, and the same continuation budget buys roughly eighty-plus segments on DeepSeek V4 Flash. We ran the cost math against what continuation actually costs; the one-line answer is that Fable 5 is a “pivotal chapters only” tool unless money is not a variable for you.

July 24: deepseek-chat retires — the easiest migration of the summer

DeepSeek’s official announcement is unambiguous: deepseek-chat and deepseek-reasoner will be fully retired and inaccessible after July 24, 2026, 15:59 UTC. Both names currently route to deepseek-v4-flash (non-thinking and thinking modes respectively), and have done so since the V4 preview shipped in late April. So the model behind your calls does not change on July 24 — only the alias stops resolving.

The migration is a one-line diff: set model to deepseek-v4-flash (add the thinking parameter if you used the reasoner), keep base_url and key. No new model to evaluate, no style-drift risk, because the weights are the ones you are already on. The only people who get burned are the ones with the old string hardcoded and nobody watching the error logs. Grep now; it takes five minutes.

July 30: GitHub Models retires entirely

The official GitHub changelog confirms it: on July 30, GitHub Models shuts down for everyone — playground, model catalog, inference API, and BYOK endpoints, including existing customers with active usage. A scheduled brownout already ran on July 16 and another lands July 23. If your script errored on those days, that was the deadline introducing itself.

The people this actually hits are the ones who used it as a free tap for trying models. With the tap gone, the next stop is going direct — and entry-tier direct pricing is genuinely small. DeepSeek V4 Flash charges $0.14 per million input tokens; for fiction-scale usage that is a few-dollars-a-month line item, not a budget decision.

August 31: Moonshot clears house — V1 classics and k2.5 go down together

Kimi K3 shipped on July 16, and with it Moonshot marked the entire moonshot-v1 series and kimi-k2.5 on its official model list as closed to new users with a full platform sunset on August 31. The older K2 previews were already discontinued on May 25. A top-up promotion returns 10-30% in bonus vouchers through August 11 (the 30% tier starts at ¥5,000). The intent is plain: everyone onto k2.6, k2.7-Code, or k3.

If you write fiction on moonshot-v1, we have benchmark data for the replacement call. Staying with Kimi means kimi-k2.6: in our nine-model palace-intrigue benchmark it placed 5th-7th, and both blind reviewers independently noted it was the only system that improved as the run went on — it opened in a shouty webnovel register and settled into period decorum. Its one recorded weakness, a mild self-copying habit on long runs, is documented in our long-run failure study. The brand-new k3 we have not tested; do not assume newer means better for continuation. Switching vendors, pick by genre: DeepSeek V4 Flash won our fantasy run, DeepSeek V4 Pro took the top tier on ornate historical fiction. Full tables in the model-picking answer page.

August 31: Sonnet 5’s introductory price expires — two bills to add up

Per Anthropic’s announcement, Sonnet 5’s introductory pricing of $2 per million input tokens and $10 output runs through August 31; from September 1 it moves to the standard $3/$15. That reads as a 50% increase, but budget for the second line: Sonnet 5 uses a new tokenizer that maps the same text to roughly 1.0-1.35× as many tokens, per Anthropic’s own footnote. The introductory price was set to make the transition “roughly cost-neutral” against that inflation — once the window closes, anyone who migrated from Sonnet 4.6 pays on both ends. If Sonnet 5 writes your long-form, redo the monthly math before September.

Why do models just disappear?

Lined up on one page, the reasons sort themselves. Product generational churn (Moonshot clearing the stage for K3). Strategic retreat (GitHub retiring a whole product). Pricing redesign (Gemini’s free tier, DeepSeek’s peak hours). And one category nobody plans around: on June 12, three days after launch, the US government applied export controls to Claude Fable 5, and because Anthropic could not verify user nationality in real time, it suspended the model for every user worldwide — for 19 days, until July 1 (Anthropic’s own account). Announced dates are not even a reliable floor: Google’s Imagen 4 shutdown was announced for August 17, and when we wired up a test integration in mid-July it was already refusing new API users.

None of this is an accusation — providers are entitled to manage their catalogs. The structural problem sits on the user’s side: a workflow welded to one vendor’s one model name is one changelog entry away from a standstill. Foreverse has been BYOK and multi-provider from day one — 60+ providers preconfigured, any OpenAI- or Anthropic-compatible endpoint addable, models switchable per continuation segment. Not because we predicted whose calendar would blow up, but because your book’s progress should not depend on anyone’s calendar at all.

How to use this page

Bookmark it; don’t memorize it. We update this page as dates land, announcements change, or new sunsets get posted, and bump the date under the title each time. Only two items need action this week: grep your code for deepseek-chat (before July 24), and check whether your Kimi calls still point at moonshot-v1 (before August 31). Everything else can wait for its date.