Opus 4.8 fast mode starts returning errors on November 30. Opus 5.5 fast costs 20% less per token, but Opus 5.5 can't skip thinking.
You have 15 days. The obvious move is cheaper per token, and mostly cheaper per request too, but read the thinking section before you trust the 20%.

Photo by Scott Parry on Unsplash
- After November 30,
speed: "fast"on Opus 4.8 is an error, not a slower answer. Change the model ID before then. - Opus 5.5 fast cut our test turn from $0.195 to $0.138, as long as its always-on thinking doesn't push output up 71% or more. For hard agent tasks, Fable 5.2 fast at medium effort is cheaper per task than Opus 5.5 fast at xhigh.
On October 30 Anthropic put a line in its release notes that a lot of Claude Code users scrolled past: fast mode for Claude Opus 4.8 is deprecated, with removal on November 30, 2026. After that date, a request to claude-opus-4-8 with speed: "fast" returns an error.
That last part is the bit to act on. When Opus 4.6 lost fast mode in June, its fast requests quietly ran at standard speed and standard price. Opus 4.8 doesn't get that grace. Anything still sending the fast flag on December 1 breaks.
The model itself isn't going anywhere yet. The deprecations page still lists claude-opus-4-8 as active, with retirement no sooner than May 28, 2027. Only the fast lane closes.
Where fast mode still exists
| Model, fast mode | Input | Cache hit | Output | Status |
|---|---|---|---|---|
| Opus 4.8 | $10 | $1.00 | $50 | Errors from Nov 30 |
| Opus 5 | $10 | $1.00 | $50 | Research preview, waitlist |
| Opus 5.5 | $8 | $0.40 | $40 | Generally available |
| Fable 5.2 | $16 | $0.40 | $80 | Available since Nov 9 |
Per 1M tokens, from Anthropic's fast mode docs and pricing page, November 15, 2026. Anthropic doesn't print fast cache rates; we applied each model's standard cache multiplier (0.1x, 0.05x, 0.025x) to its fast input price, as the pricing page says to.
Every fast card is exactly twice the model's standard card. Anthropic claims up to 2.5x the output tokens per second, and says plainly that time to first token doesn't improve. Fast mode doesn't work with the Batch API on any model, and Sonnet and Haiku don't have it at all.
Opus 5 is the other $10/$50 option, the same price you pay today, but it's still behind a waitlist. For most teams that leaves two real choices: Opus 5.5 at $8/$40 or Fable 5.2 at $16/$80. Opus 5.5 is also where Anthropic points you, and it's the default for /fast in Claude Code from v2.1.280.
One agent turn, four ways
A rate card doesn't tell you much on its own, so here's a typical coding-agent turn: a 50K-token context of which 45K is a cache hit, 5K is new, and 2K tokens come back.
| Model | Per turn | 100K turns a month |
|---|---|---|
| Opus 4.8 fast (today) | $0.195 | $19,500 |
| Opus 5.5 fast | $0.138 | $13,800 |
| Fable 5.2 fast | $0.258 | $25,800 |
| Opus 4.8 standard | $0.0975 | $9,750 |
| Opus 5.5 standard | $0.069 | $6,900 |
Our arithmetic at list prices, same token counts on every row.
Opus 5.5 fast comes out 29% cheaper than what you pay now, more than the headline 20%. The extra comes from the cache hit. Opus 5.5 bills cache reads at 0.05x input where Opus 4.8 bills 0.1x, so the 45K cached tokens cost $0.018 instead of $0.045. The more of your context sits in cache, the bigger the gap.
Fable 5.2 fast costs about a third more than you pay today on this turn. That's the price of the strongest model Anthropic sells, and we'll come back to whether it's worth it.
The catch: Opus 5.5 always thinks
That table assumes both models write the same 2K tokens. They might not. Anthropic's migration guide says every Opus 5.5 request runs adaptive thinking, and it can't be disabled. If your Opus 4.8 calls never set thinking, they ran without it. After the switch they'll think, and thinking bills as output, at $40 per million in fast mode.
On our example turn, Opus 5.5 fast stays cheaper until it writes about 3,425 output tokens, 71% more than 4.8's 2K. Past that, you're paying more for the same work. We can't tell you where your traffic lands, because it depends on the task and the effort setting. Opus 5.5 defaults to medium effort, so leave it there to start and set effort: "low" on short, mechanical calls.
Run a day of real traffic on both and compare usage.output_tokens before you move everything. If your 4.8 calls already used extended thinking, this mostly doesn't apply and the 29% is closer to real.
When Fable 5.2 fast is the cheaper task
Per token, Fable 5.2 fast is exactly twice Opus 5.5 fast. Per task it can be the other way round, because Fable at medium effort gets a near-identical score with far fewer tokens than Opus at xhigh.
| Model and effort | AA score | Standard $/task | Fast $/task | Minutes (standard) |
|---|---|---|---|---|
| Fable 5.2 medium | 56.4 | $3.80 | $7.60 | 7.4 |
| Opus 5.5 xhigh | 57.4 | $6.05 | $12.10 | 15.8 |
| Fable 5.2 high | 61.8 | $9.08 | $18.16 | 15.5 |
| Opus 5.5 max | 59.1 | $10.04 | $20.08 | 25.4 |
Standard cost and time from Artificial Analysis Intelligence Index v5 (Opus 5.5, Fable 5.2). Fast cost is our arithmetic: fast mode doubles every line of the card, and AA doesn't benchmark fast endpoints.
Fable 5.2 at medium scores 56.4 for $7.60 a task in fast mode. Opus 5.5 at xhigh scores 57.4 for $12.10. That's one point for 59% more money, and at standard speed Opus xhigh also takes about twice as long per task, 15.8 minutes against 7.4. If you turn on fast mode because you want answers sooner, spending half the tokens is the bigger speedup.
Opus 4.8 isn't on AA's current index, so we can't put it in this table. Our Fable 5.2 effort post has the full effort ladder if you want to pick a different level.
Three things that cost money during the switch
Caches don't carry over. A new model means a cold cache, and the fast mode docs add that fast and standard requests don't share cached prefixes even on the same model. Writing a 50K context to the 5-minute cache on Opus 5.5 fast costs $0.50, about 3.6 normal turns. That's fine once per session. It adds up if your code flips speeds mid-conversation.
In Claude Code, fast mode is billed to usage credits on every subscription plan, even if you have plan usage left, according to the Claude Code docs. Switching with /model to Opus 5.5 needs v2.1.280 or later, and fast stays on after the switch. Or run /fast off to stay on Opus 4.8 at standard price.
Priority Tier isn't supported on Opus 5.5. Fast mode doesn't work with Priority Tier either, but if some of your Opus 4.8 traffic uses Priority Tier and some uses fast, the Priority part has to stay on 4.8 or find another home.
Before December 1
Grep for speed and fast-mode-2026-02-01 this week, not on November 29. Opus 5.5 doesn't need the beta header, so you can drop it as you change the model ID.
For most fast traffic, Opus 5.5 fast is the right landing spot. It's cheaper on every line of the card and two generations newer, and the thinking overhead has to reach 71% before it eats the saving on a cache-heavy turn like ours.
If you were running 4.8 fast at high effort on hard agent tasks, test Fable 5.2 fast at medium next to Opus 5.5 at xhigh. On AA's numbers it's cheaper and quicker for nearly the same score. And if you only turned fast on because it was the default, ask whether you need it: Opus 5.5 at standard speed is $0.069 on our turn, about a third of what you pay now.
See every Claude rate side by side on the pricing table, check Opus 5.5's model page, or run your own token counts through the calculator. Our Fable 5.2 pricing post has the full Fable card. For other end-of-month deadlines, see our November price changes roundup.
Sources
- Claude Developer Platform release notes - October 30 deprecation, removal date, error behaviour, Opus 4.6 fallback precedent
- Fast mode docs - supported models, prices, speed claim, cache isolation, beta header, Batch exclusion
- Claude pricing - standard and fast rates, cache multipliers
- Opus 5.5 migration guide - always-on adaptive thinking, Priority Tier, cache minimum
- Model deprecations - Opus 4.8 status and earliest retirement date
- Claude Code fast mode docs - deprecation warning, /fast default, usage credits, version requirements
- Artificial Analysis model pages - index scores, cost per task and time per task