The Great Model Purge: 13 Deprecations in One Week as the GPT-3.5 Era Ends

For an industry addicted to launch events, the week of August 10–16, 2026 delivered something unusual: a frontier AI tracker recorded zero new model releases and zero price changes. What it did record were thirteen deprecation updates — a coordinated wave of model retirements across OpenAI, Anthropic, and Google that reads less like routine housekeeping and more like the closing chapter of an era. The remnants of the GPT-3.5 generation, the model family that arguably started the modern AI boom, now have a formal end date.

What actually happened

According to AI Flash Report’s model tracker for Week 33 of 2026, the industry’s model fleet shrank instead of growing. Four models were shut down outright during the week, and nine more had shutdown dates formally announced.

The completed shutdowns:

• claude-opus-4-1-20250805 (Anthropic) — shut down August 5, 2026, replaced by claude-opus-4-8. Anthropic had notified developers on June 5, 2026, giving roughly a two-month migration window.
• embedding-2-preview (Google) — shut down August 10, 2026, replaced by gemini-embedding-2.
• gpt-5.2-chat-latest (OpenAI) — shut down August 10, 2026, replaced by gpt-5.6-sol.
• gpt-5.3-chat-latest (OpenAI) — shut down August 10, 2026, also replaced by gpt-5.6-sol.

The newly announced shutdown dates are where the historical significance lies. On September 28, 2026, OpenAI will pull the plug on babbage-002, davinci-002, gpt-3.5-turbo-1106, and gpt-3.5-turbo-instruct — the last of the completion-style and GPT-3.5-era API surfaces, all mapped to gpt-5.6-terra as the successor. Then on October 23, 2026, the fine-tuned derivatives follow: ft-babbage-002, ft-davinci-002, ft-gpt-3.5-turbo, ft-gpt-4, and ft-gpt-4.1-nano-2025-04-14, redirected to gpt-5.6-terra and gpt-5.6-luna respectively.

If your application still calls gpt-3.5-turbo-instruct, this is the quarter to fix that.

Not a surprise, but still a milestone

None of this happened without warning. OpenAI’s April 2026 deprecation notice stated plainly that “over the next three to six months, we will deprecate a set of older OpenAI models to improve reliability and simplify model selection.” The February 13, 2026 retirement of GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini already cleared the middle generation. This week’s announcements finish the job at the bottom of the stack.

The GPT-5.6 family — Sol, Terra, and Luna, launched July 9, 2026 — is the destination for nearly every migration path. OpenAI has structured it as a clean three-tier ladder: Sol as the flagship at $5 per million input tokens and $30 per million output tokens, Terra as the balanced mid-tier at $2.50/$15, and Luna as the fast budget option at $1 input. The pattern in the deprecation table is deliberate: conversational legacies route to Sol, legacy base models and instruct variants route to Terra, and the smallest fine-tunes route to Luna.

The bigger picture: lifecycles are collapsing

The most important data point isn’t any single shutdown — it’s the rate. Analysis from deprecation trackers suggests model lifecycles have compressed to 6–12 months at the aggressive end, with OpenAI turning over its fleet measurably faster than Anthropic, which tends to grant longer tail windows to its Claude lines. Google sits somewhere in between, and its retirement of a mere embedding preview this week suggests its cleanup is more targeted than structural.

There’s also a structural break coming that dwarfs the model list: the Assistants API shuts down on August 26, 2026, ten days from now. For developers, that’s a bigger migration project than any individual model swap — the Responses API is not a drop-in replacement, and applications built on assistants, threads, and runs need real re-architecture, not a model-string edit.

Why providers are doing this

Fleet consolidation isn’t just aesthetics. Every legacy model snapshot kept alive consumes serving capacity, safety-evaluation surface area, and support burden. In a year when inference demand is straining datacenter buildouts, retiring a GPT-3.5-era model that a handful of legacy integrations still call is cheap capacity recovery. There’s also a security angle: older models lack the safety classifiers and watermarking that newer releases carry — Anthropic, for instance, confirmed that Claude models launched on or after August 2, 2026 weave an invisible machine-readable watermark into generated text. Every model retired is a model that no longer bypasses that infrastructure.

What developers should do now

Three practical takeaways from this week’s purge:

1. Audit your model strings this week. Anything matching gpt-3.5*, babbage*, davinci*, or ft- prefixed variants is now on a dated countdown. The September 28 and October 23 deadlines arrive faster than a sprint cycle.
2. Don’t wait for the redirect. Although OpenAI maps deprecated models to successors, behavior and pricing differ — Terra is not a GPT-3.5-priced model, and evals will shift. Treat migration as a testing project, not a find-and-replace.
3. Check the Assistants API clock. August 26 is the hard sunset. If you’re still on it, that migration should be outranking everything else on your backlog.

The week without a launch was, in its own way, the loudest signal of the year: the industry has decided that the future belongs to consolidated, watermarked, three-tier model families — and the past gets thirty days’ notice. The era when a model launched in 2022 could live forever on the API is officially over.