Week 29, 2026: OpenAI legacy shutdown in 7 days, Anthropic memory API change in 6
The rule of this digest: if it doesn't break your build, change your bill, or carry a deadline - it's not here. About the radar →
DEPRECATIONOpenAI: first shutdown wave of legacy snapshots - July 23
Announced 2026-04-22, the first of two shutdown waves hits in 7 days: legacy gpt-3.5-turbo and gpt-4 variants, o1 and o3-mini snapshots - including their fine-tuned versions. Second wave: October 23.
Do: grep your codebase for these model IDs today; anything still pointing at them starts erroring next Thursday. OpenAI's deprecations page has the per-model wave assignment - verify yours there.
Source: OpenAI deprecations
BREAKINGAnthropic: memory-listing semantics change - old header inherits it July 22
New beta header agent-memory-2026-07-22 changes GET /v1/memory_stores/{id}/memories: fixed server-side ordering (order_by/order ignored), depth only 0/1/absent (else 400), path_prefix must end with / and match whole segments; old page cursors invalid. On July 22 the existing managed-agents-2026-04-01 header adopts the same behavior - and current SDKs (Python 0.116.0, TS 0.110.0+) already send the new header.
Do: if you use Claude Managed Agents / memory stores, adapt list-handling code (sorting, pagination, path_prefix) and pin/bump SDKs deliberately before Wednesday, July 22.
Source: Anthropic API release notes
PRICEOpenAI GPT-5.6: cache writes are no longer free
GPT-5.6 family launched July 9 (sol $5/$30, terra $2.50/$15, luna $1/$6 per 1M tokens). Cache reads stay at 10% of input price - but explicit cache writes are now billed at 1.25× input (sol: $6.25/1M). If your production workload leans on prompt caching with explicit writes, the math changed.
Do: re-run your cache economics before migrating to 5.6; heavy-write patterns may cost more than on 5.5 despite lower headline prices.
Source: OpenAI pricing
DEPRECATIONAnthropic: fast mode on Opus 4.7 errors after July 24
Two different behaviors, easy to confuse: on Opus 4.6, speed: "fast" already silently degrades to standard mode (billed standard - check usage.speed). On Opus 4.7, the same parameter returns an error after July 24.
Do: move fast-mode workloads to Opus 4.8; audit responses' usage.speed if you assumed 4.6 fast mode still works.
Source: Anthropic API release notes
PRICEReminder with a date: Sonnet 5 intro pricing ends Aug 31
Claude Sonnet 5 launched June 30 at an introductory $2/$10 per MTok. From September 1 it's $3/$15 - a +50% jump on inputs. Also note: Sonnet 5's new tokenizer produces ~30% more tokens for the same text vs Sonnet 4.6, so per-request cost rises more than the price table suggests.
Do: budget Sonnet 5 workloads at $3/$15 × ~1.3 token factor now, not on September 1.
Source: Anthropic API release notes
DEPRECATIONAWS Bedrock: Agents Classic closes to new customers July 30; Claude legacy EOLs stack up
Bedrock Agents (Nov 2023) is now "Agents Classic": from July 30 no new customers (allowlist = accounts active in the last 12 months), frozen model catalog; AWS points to AgentCore. Separately: Claude Opus 4.1 on Bedrock went Legacy (EOL 2027-01-08, possible price increases from 2026-10-08), and four Claude 3.x Sonnet variants hit EOL July 30.
Do: new agent projects on AgentCore only; check your Bedrock model IDs against the lifecycle page this week.
Sources: Agents Classic notice · Bedrock model lifecycle
PRICEBedrock: Claude Sonnet 4 entered "extended access" - price can rise anytime
Since July 14, anthropic.claude-sonnet-4-20250514-v1:0 on Bedrock runs in public extended access until EOL 2026-10-14 - usable, but pricing "may increase, set by the model provider". On the first-party Claude API this model was already shut down June 15.
Do: migrate remaining Sonnet 4 traffic to Sonnet 4.6 or Sonnet 5 now; don't wait for the surprise line item.
Source: Bedrock model lifecycle
POLICYAnthropic: Fable 5 / Mythos 5 access restored
After a brief suspension at the turn of June/July, API access to claude-fable-5 and claude-mythos-5 was restored July 1.
Do: if you wired fallbacks or model-swaps during the outage, you can unwind them - and keep the fallback code path; this incident proved you need one.
Source: Anthropic API release notes
Worth knowing (the only NEW items that change decisions)
- Anthropic API keys can now expire (Console-set, with pre-expiry email). If your org has key-rotation policies, this is the native mechanism. Source
- GPT-5.6 is GA on AWS Bedrock with 90%-discount cached input - a multi-cloud option if you need GPT with AWS-side data residency. Source
Filtered out this week: 10 NEW items (realtime model updates, console redesigns, admin betas, developer logs). They're launches, not obligations.
Next 30 days - from the deprecation calendar
- Jul 21 - Anthropic retires
claude-mythos-preview - Jul 22 - Anthropic memory-listing behavior switch (above)
- Jul 23 - OpenAI legacy snapshot shutdown wave 1 (above); Azure retires gpt-realtime-mini (2025-10-06)
- Jul 24 - Opus 4.7 fast mode starts erroring (above)
- Jul 30 - Bedrock: Agents Classic gate + Claude 3.x Sonnet EOLs (above)
- Jul 31 - Mistral retires Small 3.2, Devstral 2, Magistral 1.2, Nemo 12B; Azure retires Stable Diffusion 3.5 models
- Aug 5 - Anthropic retires
claude-opus-4-1on the Claude API - Aug 10 - OpenAI retires
gpt-5.2-chat-latest,gpt-5.3-chat-latest - Aug 13 - Azure Foundry retires DeepSeek-R1 (Legacy)
- Aug 15 - Azure Foundry retires MAI-Image-2e / MAI-Image-2 previews