Fable 5.2 Agentic Coding: Effort, Fast Mode, Call Today

The Fable 5.2 migration is mostly a string change, until it is a 400. On SandBase the string is anthropic/claude-fable-5.2.

I read the migration guide because the launch post makes 5.2 sound like a drop-in that is 20% cheaper than Fable 5.1. The guide agrees the surface is close, then lists the breaks: thinking cannot be disabled, forced tool choice returns 400, sampling parameters return 400, and a workspace without 30-day retention returns 400. Effort, if you omit it, is high, one level above Opus 5.5’s medium. Fast mode is optional and doubles the price.

That is the Claude API. SandBase, checked 10 November 2026, lists Claude Fable 5.2. Rehearse the hard loop on that id, and keep Claude Opus 5.5 as the daily driver. Claude Fable 5.1 is the previous Fable id when you already have a score on it. The release note is what shipped. The cost and routing table is when not to bother.

Key takeaway

• Fable 5.2 supports five effort levels: low, medium, high, xhigh, max. The API default is high. Anthropic says start there, step to xhigh only where you measured a gain, and reserve max for a ceiling run.
• Adaptive thinking is always on. thinking: {type: "disabled"} returns 400. Thinking tokens count toward max_tokens and are billed as output.
• tool_choice of any or tool returns 400. auto and none still work. Omit temperature, top_p, and top_k.
• Fast mode is speed: "fast", up to about 2.5× token generation, at $16 / $80 per million tokens. The announcement lists Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry. It does not list Bedrock in that sentence.
• SandBase id: anthropic/claude-fable-5.2. The card shows $8 / $40, 1M context, and 128K max output, with no discount on that row. Do not send the hyphenated claude-fable-5-2. The card’s Chat Completions schema does not list effort or speed, and I have not verified that a live call forwards them.

Where to read before you change a client

The parameter list is useless if you pick the wrong job. When to leave Opus is the routing page. What the 9 November launch confirmed is the rate card and the live SandBase card.

The loop you are probably already running is described in Fable 5.1 long-running agents and in why a long agent forgets the goal. Those pages are about 5.1. The recovery rules (a task record, a regression check, a bounded restart) do not expire because a new id showed up. Mythos versus Fable access still applies: the public agent model is Fable, not Mythos. Opus 5.5 pricing is the bill if you stay on the daily driver. The October rumor is the leak timeline, not a setup guide.

Effort is the cost knob

Anthropic’s effort doc, retrieved with the overview on 10 November 2026, says Fable 5.2 supports all five levels and that the API default is high. Most of the gain they describe is already at high. Above that, gains are uneven. xhigh is for long-running coding where you have measured a difference. max optimizes for capability without regard to token cost, and they say it is often only a small step past xhigh while being slower. medium and low are for slices where your evals hold.

Claude apps and Claude Code are a different default. The announcement says Medium effort is the default there, and that at Medium, Fable 5.2 used fewer tokens than Fable 5.1 on hard tasks in early testing. If you copy a Claude Code session into an API client and forget effort, you silently move from the app’s Medium to the API’s high. That is the kind of bug that shows up as a bill, not as an error.

Effort is a behavioral signal, not a token cap. At high and above, set max_tokens large. It is a hard limit on thinking plus text. A small cap can stop on max_tokens after a thinking block and before any text. Prompts that used to say “answer directly, don’t deliberate” do not, in Anthropic’s testing, stop 5.2 from thinking at medium and above. Lower the effort instead of pleading with the prompt.

On the Claude API the field is output_config.effort. This article does not show that field in the SandBase sample. I did not send a request that proves the gateway accepts it.

Fast mode is a second price, not a smarter model

Fast mode on Fable 5.2 is the same model with a higher serving rate. Anthropic’s number is up to 2.5× token generation, at $16 per million input tokens and $80 per million output tokens. The migration guide names speed: "fast" on Messages API requests. Opus 5.5’s fast mode, on that same guide, is $8 / $40, which is Fable 5.2’s ordinary price. So “turn on fast” on Opus lands you on the sticker Fable charges for normal speed, and “turn on fast” on Fable lands you at double that.

I would not enable it for a coding agent by default. Tool rounds, test runs, and checkout time do not get 2.5× faster because tokens do. Turn it on when a trace shows the wait inside generation, the session is on the critical path, and you have already cut effort as far as quality allows. Mythos fast mode, the announcement says, is only for organizations in the verification programs.

The platforms named for Fable fast mode are Claude Code, the Claude Platform, Claude on Google Cloud (both hosting options), and Microsoft Foundry (Anthropic hosted). I am not adding Bedrock to that list. The overview’s platform list for the model itself does include Bedrock. Availability of the model and availability of fast mode are different sentences.

What returns 400

Fable 5.2 overview, Good to know, captured 10 November 2026. Adaptive thinking cannot be disabled. tool_choice types any and tool return 400. Omit temperature, top_p, and top_k. The model requires 30-day retention and is not on zero data retention unless Anthropic authorizes it. A linked card says editing earlier turns invalidates thinking blocks, for every account.

The same overview, specifications. Use these ids on Anthropic’s surfaces. They are not SandBase model strings.

The breaks I would fix before a launch-day switch, all from the migration guide:

1. Retention. Fable 5.2 requires 30-day retention. A Claude API request from a workspace without it returns 400 invalid_request_error. Opus 5.5 can sit on a zero-data-retention arrangement; this model does not, unless Anthropic expressly authorizes it. The announcement describes Enterprise Frontier Safeguards, in private beta, as the path that runs safety checks on data stored in the customer’s infrastructure. Eligible customers can keep using the model with zero data retention before that monitoring is in place. If you are not in that conversation with Anthropic, assume 30 days. Do not discover this in production.

2. Tool choice. {type: "any"} and {type: "tool", name: "..."} return 400, including on the token-counting endpoint. auto and none work. The guide’s replacement is auto, the tool named in the instruction, and strict: true where your account can use strict tools. Customer-managed encryption keys are called out as a case where strict tools and JSON outputs are unavailable for this model. Then you validate the tool input yourself.

3. Sampling. Omit temperature, top_p, and top_k. Sending the usual values 400s, same as on Fable 5.1.

4. Thinking blocks. They come first. Code that reads content[0].text will grab the wrong block. In a tool loop, send thinking blocks back unchanged. Preserved thinking now applies to every account: editing an earlier turn invalidates those blocks. The guide also says Fable 5.1 cannot read Fable 5.2’s thinking blocks, so a mixed history is not a free upgrade.

5. Subagents at high effort. Anthropic says that at xhigh and max, in an agentic coding product, 5.2 started review subagents that 5.1 did not spawn on the same tasks. That can dominate the bill. If your harness treats every child session as free, it is wrong.

6. One older computer-use tool. computer_20251124 returns 400 on every platform except Amazon Bedrock. If a harness still pins that tool id, the failure looks like a model bug. It is a version pin.

Confirmed vs unknown

Try today on SandBase

The Anthropic catalog on SandBase, 10 November 2026. The Claude Fable 5.2 tile is Available, id anthropic/claude-fable-5.2, $8.00 / $40.00 per million tokens, 1,000,000 context tokens. In the same crop, Claude Fable 5.5 is badged Available, with a coming-soon chip and Contact for pricing. That tile is not a 5.5 calling path, and this article does not treat 5.5 as live.

The id to put in today’s Fable call: anthropic/claude-fable-5.2. Displayed price $8.00 / $40.00 per million tokens, cache read $0.20, cache write $10.00, context 1M, max output 128K, status Available. No discount on that row. Captured with the release note.

Two curls. Same Chat Completions route the card documents. No output_config, no speed. The card’s schema lists temperature and does not list effort or speed. I am not sending Claude-only fields I have not seen accepted. If you need effort or speed, you are on the Claude API until you have checked them here.

curl --fail-with-body --silent \
  -X POST "https://api.sandbase.ai/v1/chat/completions" \
  -H "Authorization: Bearer $SANDBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-fable-5.2",
    "messages": [
      {"role": "user", "content": "You are editing a billing client used by two repos. State the invariant you will not break, the test that proves it, and the stop condition if the test cannot be run."}
    ],
    "max_tokens": 4096
  }'

The daily-driver comparison uses anthropic/claude-opus-5.5. Same morning: $3.40 input and $17 output per million tokens, 15% off list $4 / $20. This is not the 5.2 card.

curl --fail-with-body --silent \
  -X POST "https://api.sandbase.ai/v1/chat/completions" \
  -H "Authorization: Bearer $SANDBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-opus-5.5",
    "messages": [
      {"role": "user", "content": "You are editing a billing client used by two repos. State the invariant you will not break, the test that proves it, and the stop condition if the test cannot be run."}
    ],
    "max_tokens": 4096
  }'

Run both on a task you already score. Record pass, tokens, and dollars from the response usage, not from this paragraph. Swap in anthropic/claude-fable-5.1 only when you want the previous Fable baseline. The long-running agent writeup is one way to keep that score honest across retries. I did not attach a fresh completion to this page.

On the Claude API the documented shape is different: model claude-fable-5-2, max_tokens large enough for thinking, output_config.effort set on purpose, no temperature, tool_choice left on auto unless you mean none. Add speed only after a trace says generation is the wait. That paragraph is a checklist from their docs. It is not the SandBase request above, and it is not a transcript from a run I made.

The SandBase card is already HTTP 200. Re-read its schema before you copy Claude fields across. The routing page is how you decide the job deserved 5.2 rather than Opus.

FAQ

What effort should a coding agent use?

On the Claude API, Anthropic says start Fable 5.2 at high, which is also the default. Move to xhigh only after your evals show a gain on long jobs. Use max to measure the ceiling, not as a production default. Opus 5.5 defaults to medium, so do not copy an omitted effort from one client to the other and expect the same spend.

Is fast mode worth it for agents?

Sometimes, for the turns that are waiting on tokens. It is up to about 2.5× generation at 2× price ($16 / $80). It does not speed up your test suite. Leave it off until a trace says otherwise.

Why did tool_choice start returning 400?

Fable 5.2 rejects forced tool choice (any and tool). Keep auto, name the tool in the instruction, and use strict tools only if your account supports them. none still works for a turn that must not call tools.

Can I point SandBase at claude-fable-5-2 today?

Use the dotted id. SandBase’s string is anthropic/claude-fable-5.2. The hyphenated claude-fable-5-2 is the Claude API id. The card was HTTP 200 on 10 November 2026. The release page has the screenshot. Earlier the same day that path was 404. That is not the current card.

Do I need Mythos for coding agents?

No, unless the work is in the verified cyber or life-sciences programs. Anthropic says Fable and Mythos are the same underlying model, and Fable is the one with the safeguards, offered generally. See access tiers for 5.1, which is the same split in the previous generation.

Related reading

• Fable 5.2 release: price, Mythos, and the live card.
• Fable 5.2 vs Opus 5.5 vs GPT-6 Astra.
• Fable 5.1 long-running agents and the closed-loop failure mode.
• The leak timeline, so a 5.5 rumor does not get mixed into this client change.

Ship the rehearsal, not the rumor

Keep one config key for the model id. Point the ordinary loop at Opus 5.5. Point the job that already needed Fable at anthropic/claude-fable-5.2. Leave Fable 5.1 in the config only as the scored baseline. Log effort in a comment next to that key, because the card’s Chat Completions schema does not list the field. Fix tool choice and retention on the Claude API before you celebrate the lower list price.

Sign in and open Claude Fable 5.2. The Anthropic catalog shows that tile as Available at $8 / $40. Send the invariant prompt above and keep the usage object.

Docs retrieved 10 November 2026 (Asia/Shanghai): Fable 5.2 overview, migration guide, effort guide, the 9 November announcement, and the SandBase cards linked above. No Fable 5.2 completion was captured for this article.