🌏 中文版
What changed
OpenAI's deprecations page has turned the notice it sent on June 11, 2026 into a firm shutdown date: December 11, 2026. gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, and o3-pro are all going dark, right at the six-month minimum notice period for generally available models. This isn't one model retiring — it's the entire GPT-5 and o3 lineup closing out at once. OpenAI names a specific successor for each: gpt-5/o3 moves to gpt-6.1-sol, gpt-5-mini moves to gpt-6.1-terra, gpt-5-nano moves to gpt-6-luna, and gpt-5-pro/o3-pro moves to gpt-6.1-sol with the Responses API's reasoning.mode: pro flag instead of a standalone pro model. Lining up the six old-to-new rate pairs reveals something uneven: the flagship tier (gpt-5→gpt-6.1-sol) and the pro tier (gpt-5-pro/o3-pro→the same gpt-6.1-sol) get steep rate cuts, while the mini tier (gpt-5-mini→gpt-6.1-terra) gets a steep rate increase — the same announcement is good news and bad news depending on which budget tier you're running on.
Before and after
| Item | Before (until 2026-12-11) | After (recommended replacement) | Change | Effective |
|---|---|---|---|---|
| gpt-5 Input | $1.25/1M tokens | gpt-6.1-sol: $2.00/1M tokens | ↑60% | 2026-12-11 |
| gpt-5 Output | $10.00/1M tokens | gpt-6.1-sol: $10.00/1M tokens | flat | 2026-12-11 |
| gpt-5-mini Input | $0.25/1M tokens | gpt-6.1-terra: $1.00/1M tokens | ↑300% | 2026-12-11 |
| gpt-5-mini Output | $2.00/1M tokens | gpt-6.1-terra: $5.00/1M tokens | ↑150% | 2026-12-11 |
| gpt-5-nano Input | $0.05/1M tokens | gpt-6-luna: $0.10/1M tokens | ↑100% | 2026-12-11 |
| gpt-5-nano Output | $0.40/1M tokens | gpt-6-luna: $0.50/1M tokens | ↑25% | 2026-12-11 |
| gpt-5-pro Input | $15.00/1M tokens | gpt-6.1-sol (reasoning.mode: pro): $2.00/1M tokens | ↓87% | 2026-12-11 |
| gpt-5-pro Output | $120.00/1M tokens | gpt-6.1-sol (reasoning.mode: pro): $10.00/1M tokens | ↓92% | 2026-12-11 |
| o3 Input | $2.00/1M tokens | gpt-6.1-sol: $2.00/1M tokens | flat | 2026-12-11 |
| o3 Output | $8.00/1M tokens | gpt-6.1-sol: $10.00/1M tokens | ↑25% | 2026-12-11 |
| o3-pro Input | $20.00/1M tokens | gpt-6.1-sol (reasoning.mode: pro): $2.00/1M tokens | ↓90% | 2026-12-11 |
| o3-pro Output | $80.00/1M tokens | gpt-6.1-sol (reasoning.mode: pro): $10.00/1M tokens | ↓87.5% | 2026-12-11 |
The pro-tier cuts are a rate comparison, not a billing guarantee. reasoning.mode: pro turns what used to be a separately priced Pro model into extra thinking tokens on the same model. If your migrated workload needs a longer reasoning chain, it may burn through more tokens than before, so comparing per-token rates alone understates the real cost — test against your own tasks before counting on the full discount.
Cost estimate
Scenario: a team running gpt-5-mini for high-volume text classification and extraction, 50,000 requests a day (averaging 800 input tokens + 200 output tokens each), over a 30-day month.
| gpt-5-mini (current) | gpt-6.1-terra (replacement from Dec 11) | Monthly change | |
|---|---|---|---|
| Input cost/month (1,200M tokens) | $300.00 | $1,200.00 | +$900.00 (↑300%) |
| Output cost/month (300M tokens) | $600.00 | $1,500.00 | +$900.00 (↑150%) |
| Total | $900.00 | $2,700.00 | +$1,800.00 (↑200%) |
This scale of classification/extraction work is exactly what the mini tier is usually used for, and the monthly bill triples. By contrast, if the same team also runs gpt-5-pro for a smaller volume of long-reasoning tasks (say, 500 requests a day, averaging 3,000 input + 2,000 output tokens), moving to gpt-6.1-sol (pro mode) at list price drops the monthly cost from $367.50 to $90.00, a 75% cut. One company, one announcement, two completely opposite outcomes depending on usage pattern.
Impact on developers and enterprises
Who benefits most
Teams running gpt-5-pro or o3-pro for deep reasoning and long-form report generation benefit the most — list prices drop more than 85%, and OpenAI has turned "Pro" from a gated, separately licensed model into a request parameter anyone can set. On the flip side, teams running gpt-5-mini for high-frequency, low-cost tasks (classification, routing, summarization/extraction) are the only losers in this announcement, and that happens to be the mini tier's single biggest use case.
Competitive impact
The "budget tier" ranking after December 11 (input / output, USD per 1M tokens):
| Model | Input | Output | Note |
|---|---|---|---|
| GPT-6 Luna / Claude Haiku 5.5 (from) | $0.10 | $0.50 | gpt-5-nano's replacement, matches Haiku 5.5's floor tier |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | not affected by this announcement |
| GPT-6.1 Terra (from Dec 11) | $1.00 | $5.00 | gpt-5-mini's replacement, now priced well outside "mini/nano" territory |
| GPT-6.1 Sol / Claude Sonnet 5.5 | $2.00 | $10.00 | gpt-5/o3's replacement, matches Sonnet 5.5 |
After the increase, gpt-6.1-terra ($1.00/$5.00) costs more than three times gemini-3.5-flash-lite ($0.30/$2.50). Teams that picked gpt-5-mini for its spot between nano-tier pricing and mid-tier quality now face a rate that's only half of the mid-tier model — the price-to-capability position that made the mini tier attractive is gone.
Recommendations
- If you're running
gpt-5-minifor high-volume, low-unit-cost tasks: benchmark bothgpt-6-luna(the nano replacement, which only rises slightly) andgemini-3.5-flash-liteagainst your own data before defaulting to OpenAI's named replacement,gpt-6.1-terra— "recommended replacement" tracks capability, not price. - If you're running
gpt-5-prooro3-pro: testreasoning.mode: proon a small slice of traffic before December 11 to measure actual thinking-token usage on your workload; don't assume an 85% bill reduction just because the list price dropped that much. - If you're running plain
gpt-5oro3: moving togpt-6.1-solis roughly neutral-to-slightly-worse (input costs more, output is about flat), so weigh the migration effort, or considergpt-6.1-terraas a downgrade option for part of your traffic — even after its price hike, it's still half the price of Sol.
Time-sensitive note
⚠️ Shutdown date: December 11, 2026. gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-pro, and every snapshot and fine-tune built on them will stop accepting requests. Migration guide: GPT-6 migration guide.
Today's takeaway
An "officially recommended replacement model" often gets read as an implicit promise that the price will be similar too — this announcement breaks that assumption. In the same table, the flagship and pro tiers drop 60%-92% while the mini tier rises 150%-300%, with no flag distinguishing which rows go up and which go down. When you're tracking pricing changes, a named replacement only tells you the vendor considers the two models comparable in capability — not in rate, especially at the low end, where it's easier for a vendor to retire an ultra-cheap old model and swap in a nominally equivalent but meaningfully pricier one. That kind of increase is easier to miss than a flagship-tier hike, because the total bill still looks like it belongs to "the cheap tier."
References
- OpenAI: Deprecations (official page — shutdown date and recommended replacements for gpt-5/gpt-5-mini/gpt-5-nano/gpt-5-pro/o3/o3-pro)
- OpenAI: API Pricing (official pricing page — current gpt-6.1-sol/gpt-6.1-terra/gpt-6-luna rates alongside legacy gpt-5/o3 rates)
- Anthropic: Models overview (official model overview — current Claude Haiku 5.5/Sonnet 5.5 rates, for the competitive table)
- This series, Nov 4: Gemini 2.5 Pro/Flash gets a Developer API sunset date
- tokencost.app: LLM API Price Changes, November 2026: The Dec 1 Cliff (2026-11-14, independent coverage of the same OpenAI December shutdown schedule)