o3 Retires August 26, 2026: What to Use Instead
The Short Answer
Two deadlines land on August 26, 2026: o3 leaves ChatGPT, and the Assistants API retires. OpenAI announced the o3 sunset on May 28, 2026, giving a 90-day window that closes in eight days from this writing.
The model swap is trivial. The Assistants API migration is the one that will hurt if you’ve left it late.
The Deadline Board
| What | Date | Replacement |
|---|---|---|
| GPT-5.2 models | Retired June 12, 2026 | GPT-5.6 family |
| GPT-4.5 (ChatGPT) | Retired June 26, 2026 | GPT-5.6 family |
| OpenAI o3 (ChatGPT) | August 26, 2026 | GPT-5.6 Sol (published) |
| Assistants API | August 26, 2026 | Responses API + Conversations API |
OpenAI’s general pattern is that models remain in ChatGPT roughly 90 days after a successor ships. Plan for that cadence rather than treating each announcement as a surprise.
Where o3 Workloads Should Actually Go
OpenAI’s published replacement is GPT-5.6 Sol. That’s the safe default, but it’s frequently the wrong economic answer — o3 was a reasoning model used for a wide spread of tasks, most of which don’t need frontier pricing.
| Model | Price /MTok | 30K/5K task | Take this if |
|---|---|---|---|
| GPT-5.6 Sol | $5 / $30 | $0.300 | Hardest reasoning; official replacement |
| GPT-5.6 Terra | $2 / $12 | $0.120 | Most o3 workloads land here |
| GPT-5.6 Luna | $0.20 / $1.20 | $0.012 | High-volume, simple classification |
| Claude Opus 5 | $5 / $25 | $0.275 | Long-horizon: 1M ctx, 128K output |
| Claude Sonnet 5 | $2 / $10 | $0.110 | Best mid-tier value; price now permanent |
| Gemini 3.7 Flash | $0.75 / $3.75 | $0.041 | Cheap + fast; intro rate to Dec 31, 2026 |
| Kimi K3 | $3 / $15 | $0.165 | Agentic work; open weights available |
Verified August 18, 2026.
The honest recommendation: don’t reflexively take Sol. Terra at $2/$12 covers the large majority of what teams were doing with o3 at 40% of Sol’s cost. Benchmark your own top three prompts against Terra first, and only escalate to Sol where Terra measurably fails.
If your o3 use was long analysis over big documents, Claude Opus 5 is the stronger destination — the 1M context and 128K maximum output ceiling are roughly double most rivals’, which matters when the deliverable is large.
If you were using o3 for cheap reasoning at volume, Gemini 3.7 Flash at $0.75/$3.75 is dramatically cheaper — though note that is a promotional rate through December 31, 2026, doubling to $1.50/$7.50 on January 1, 2027. Model your 2027 costs at the higher number.
The Assistants API Migration Is the Real Work
A model retirement is a string change. An API retirement is an architecture change.
The Assistants API’s threads-and-runs abstraction does not map one-to-one onto the Responses API. Conversation state moves to the Conversations API, and tool orchestration, file handling and run lifecycle all need reworking. Budget days, not hours.
Checklist before August 26:
- Find every caller. Grep for
assistants,threads,runsacross all repos, notebooks, cron jobs and internal tools. Shadow integrations are where teams get caught. - Port state handling first. Threads → Conversations API is the structural change; do it before touching model choice.
- Pin an explicit model. Migration is the right moment to stop relying on aliases and name the model you intend to run.
- Re-evaluate the model while you’re in there. You’re already editing the code path — this is the cheapest opportunity you’ll get to test Terra against Sol.
- Test long-running and error paths. Run lifecycle semantics differ; timeouts and retries are where behaviour diverges quietly.
The Wider Lesson: Pin Your Models
Four OpenAI retirements in three months — GPT-5.2 in June, GPT-4.5 in June, o3 and the Assistants API in August — is the actual signal. Model IDs are not stable infrastructure, and this isn’t OpenAI-specific: Google retired three Imagen 4 model IDs on August 17, 2026.
Three practices reduce the pain permanently:
- Never ship a floating alias to production. Pin explicit versioned IDs so a silent upstream swap can’t change your behaviour.
- Keep a routing layer. One place that maps task type → model makes the next migration a config edit. See how to choose an AI model in 2026.
- Keep an eval set. Twenty representative prompts with known-good outputs turn “is the replacement fine?” from a debate into a measurement.
Teams with those three in place will handle the next retirement in an afternoon. See also how to avoid AI vendor lock-in.
Last verified: August 18, 2026.