Source: raw/Claude_Haiku_vs_Sonnet_vs_Opus_vs_Fable_Clearly_Explained_use_cases.md (primary). Supporting launch-day commentary from the same creator: raw/Claude_Drops_OPUS_5_and...._Opus_5_vs_Fable_5.md, raw/Claude_SONNET_5_is_Here..._it_beats_Opus_4.8_for_cheaper.md, raw/Wait._Whaatttt..._Claude_just_dropped_FABLE_5_and_MYTHOS_5....md, raw/New_Claude_OPUS_4.7_Upgrade_Explained_for_non-technical_peeps.md. Ben AI’s seven-task head-to-head: raw/I_Tested_Opus_5.5_vs_Fable_5.1_on_7_Real_Use_Cases_Not_Even_Close.md (added 2026-10-28). Anthropic’s own answer, “Which Claude model should you use?”: raw/Which_Claude_model_should_you_use.md (added 2026-10-30). Anthropic’s November playbook for the Fable 5.2 lineup: ai-research/claude-dev-blog-which-claude-model-should-you-build-with-2026-11-11.md (added 2026-11-11).
Creator: Eliot Prince URL: https://www.youtube.com/watch?v=tRi9g3urhyE Duration: 18:36 Platform: YouTube Published: 2026-07-16 (supporting videos: Opus 4.7 2026-04-17, Fable 5 2026-06-09, Sonnet 5 2026-06-30, Opus 5 2026-07-24)
Eliot Prince’s guide to the Claude model picker, aimed at people who use Claude for office work rather than code. He made it after his fiancee, on a $20 Pro plan, used up her allowance in a few minutes because Opus 4.8 was selected instead of Sonnet 5. His model is four job levels: intern, everyday employee, senior specialist and outside consulting team. He recorded it on the July 2026 lineup (Haiku 4.5, Sonnet 5, Opus 4.8, Fable 5). This article maps his levels onto the September 2026 lineup and corrects the settings that have changed since.
Key Takeaways
- Default to Sonnet, especially on Pro. Eliot’s rule: most office work (writing, research, invoices, emails, light analysis) belongs on Sonnet. Move up only when the task needs deeper reasoning or fewer mistakes.
- Pick the model by the kind of thinking the job needs, not by how much the job matters. Haiku fetches and summarises. Sonnet carries out a task whose context and steps are already clear. Opus reasons over several documents where mistakes are costly. Fable takes a goal and works out the steps itself.
- Fable is for problems, not tasks. “Give it harder tasks… Don’t just get it to write emails for you.” Its advantage is planning: you state the outcome and it decides which connectors to read and what to do. On routine work it costs a lot and adds little.
- Plan on a big model, execute on a small one. Eliot often plans or analyses with Opus or Fable, then switches to Sonnet to carry out the steps “because the higher-end models have done all the hard work.” Anthropic now quotes the same split for Opus 5.5 and Sonnet 5.5 (see Sonnet 5.5).
- Max effort makes the model slower and more thorough, not faster. Eliot: putting Fable on max “isn’t going to make things work faster.” Use low effort on Sonnet for speed and save max for your hardest problems.
- Ignore the version numbers. He compares them to iPhone model numbers: they tell you a new release exists, not how much better it is.
- Update 2026-11-11 — Anthropic’s November rule: start on Opus 5.5 at
medium. With Fable 5.2 out, Anthropic’s playbook sends most work to Opus 5.5, down to Sonnet 5.5 or Haiku 5.5 for scoped or high-volume work, and up to Fable 5.2 only for the hardest technical projects, when Opus athighfalls short, or for long unattended runs. See the November section below. - Anthropic’s own rule: judge cost per task, not price per token. Its September 2026 video says the cheapest model at the lowest effort is not always the cheapest way to finish a task, because “a more intelligent model often finishes a task with less work.” Open-ended work goes to Fable 5.1 starting at medium effort; routine work stays on Opus or Sonnet. See the Anthropic section below.
The four levels, mapped to the current lineup
Prices are per million tokens (input / output) from this wiki’s model articles. The job-level comparisons and example jobs are Eliot’s. Matching his July 2026 models to the September 2026 releases is this article’s inference.
| Eliot’s level | Model in his video | Current model (2026-10-27) | Price | Example jobs he gives |
|---|---|---|---|---|
| Courier or one-day intern | Haiku 4.5 | Haiku 5.5 on the API since 2026-10-07; whether the claude.ai picker shows 5.5 is not recorded in this wiki | 0.50 for prompts up to 100K tokens, 2.50 above (Haiku 5.5, API) | Summarise an email thread or webpage; pull action items from a transcript; small scheduled tasks |
| Everyday employee | Sonnet 5 | Sonnet 5.5 | 10 | Turn a video transcript into a blog post or newsletter with a skill; fill invoice templates; prepare a one-pager before a sales call |
| In-house specialist or senior hire | Opus 4.8 | Opus 5.5 | 20 | Critique pricing from revenue, offers, hours and competitor data; review contracts for risks and unusual clauses (he runs a contract-analysis skill on Opus) |
| Outside consulting team (“your Deloitte team”) | Fable 5 | Fable 5.1 | 50 | Design a course from inbox, LinkedIn and customer data; clean up thousands of messy CRM records; “why aren’t I a millionaire?”, which read his subscriber list, YouTube account and invoices and returned a revenue plan |
- The Sonnet level gained the most from the 5.5 release. On paper Sonnet 5.5 lands within a point or two of Opus 5.5 at half the per-token price, and Anthropic keeps Opus for “complex work requiring careful judgment.” That supports Eliot’s Sonnet-first default.
- The Opus-versus-Fable gap has narrowed. Eliot’s Opus 5 launch video stressed that Opus 5 beat Fable 5 on the knowledge-work and business-workflow benchmarks Anthropic published, at half the price. Opus 5.5 now claims Fable-5.1-level work on most tasks (see Opus 5.5). So the Fable level is now only for the hardest long-horizon goals.
Effort and thinking settings
Eliot’s advice for the claude.ai and desktop picker, with the parts that have since changed:
- Low effort for speed, max effort rarely. He calls max the most common mistake: it “uses excessive tokens, results in long response times, and may hit your token limit.” For the fastest results he uses Sonnet on low. For the hardest problems he uses Fable on max. Sonnet 5.5’s benchmarks back up the caution: it scored lower at Max than at Xhigh on FrontierCode.
- “If you’re not sure, just stick in high mode. That’s the default.” This conflicts with this wiki’s recorded defaults. Sonnet 5.5 defaults to Medium in the Claude apps, and Fable 5.1 defaults to Medium in Claude.ai and Cowork. His own Sonnet 5 launch video also called medium the default.^[ambiguous] Either way, start at the default and change it only when the output is too shallow or too slow.
- Thinking can be switched off on some models. Eliot says to turn thinking off when Sonnet “overthinks” a simple job, which also saves usage. This no longer applies to Opus 5.5: thinking is always on, and effort is the only control.
- Adaptive thinking (Opus 4.7, April 2026). In his Opus 4.7 video he explains that “extended thinking” was renamed “adaptive thinking” and that Claude now decides per request whether to think. The trade-off he points out: Claude can think on simple requests and spend tokens you did not need to. See Opus 4.7 Best Practices and Model vs. Effort for the Claude Code version.
Plan cost: why the wrong model empties a Pro plan
- Higher-tier models use your allowance faster. Eliot’s fiancee’s Pro allowance ran out quickly on Opus 4.8 for invoices, emails and design copy. Switching her back to Sonnet fixed it.
- Fable on a subscription. In his Fable 5 launch video, Eliot said a Pro plan would “tap out really, really quickly” because Fable used about twice the usage of Opus 4.8. He recommended at least Max or a Team Premium seat. The terms that settled on 2026-07-20: Fable is included in Max and Team Premium at 50% of limits, while Pro and Team Standard use usage credits (see Fable 5).
- His per-chat cost example. At Fable 5’s API rates, he estimates a chat that fills a 1M-token context, about a quarter input and three quarters output, costs “30, 40 dollars.” That is his rough arithmetic, not a measured bill.
When the model refuses
- Fable 5 sent cyber and biology requests to a fallback model. Eliot hit this himself: asking Fable 5 to analyse its own launch post (which discusses cybersecurity and biology) triggered the safeguard and moved him to Opus 4.8.
- His workaround was to drop to Opus 5. He read the Opus 5 launch as “less restricted than Fable 5.” Since then, Opus 5.5 has launched with Fable-5.1-class safeguards, and most cybersecurity tasks now go to Opus 4.8 automatically. So switching to Opus is no longer a reliable way around a cyber flag.
Anthropic’s own answer: “Which Claude model should you use?” (2026-09)
Source: raw/Which_Claude_model_should_you_use.md — Anthropic’s official Claude YouTube channel (@claude), https://www.youtube.com/watch?v=71-8fJIGi34, uploaded 2026-09-18, 3:45. The presenter is not named. Upload date and linked resources from ai-research/anthropic-youtube-model-launch-video-descriptions-2026-10-30.md. Added 2026-10-30.
Anthropic made this short explainer after the Fable 5.1 launch and before Opus 5.5, so its examples are Fable 5.1 against Fable 5. It argues against the habit this article’s creators mostly teach, picking the cheapest model: “a more intelligent model can cost less to complete a task.”
What decides the cost of a task: model, effort, cache.
- Model. More capable models cost more per token, but “just like a college student can solve a maths problem in fewer steps than a fifth-grader, a more intelligent model often finishes a task with less work.”
- Effort is “how much reasoning the model does before it answers.” More effort costs more but “often gets a better result.”
- Cache. The model remembers nothing between requests, so every message reprocesses the whole conversation. Caching reuses what was already processed “at about a tenth of the cost.” That is why Robo Nuggets’ rule below (don’t switch models mid-chat) holds: a new model has no cache for your chat.
- The Fable example. Fable 5.1 has the same per-token price as Fable 5, but its cache reads cost 75% less, which cuts total cost by about 25% for typical work and up to about 45% for long agentic tasks. On CursorBench, Fable 5.1 at medium effort “produced a similar result to 5 at max effort for about a fifth of the cost.”
Anthropic’s routing rule.
| Kind of work | Examples Anthropic gives | Model and effort |
|---|---|---|
| Unconstrained, open-ended | Deep research, complex analysis, long-running agent work | Fable 5.1; start at medium effort and raise it as needed |
| Constrained, routine | Summarising a document, drafting an email, pulling data from a report | Stay on Opus or Sonnet; reach for Fable as needed |
| Subagents | Work “constrained to a specific task” | A cheaper model is often enough |
On routine tasks “a cheaper model can get a similar result, so you don’t pay for the capability the task never uses.” The closing rule: “consider the cost per task, not the price per token,” choose the model by the kind of task, and on more capable models “start with a lower effort and increase if needed.”
For admins: three controls, set for everyone or per role.
- Model entitlements decide which models users can pick.
- Effort caps set the highest effort users can select for a model.
- Defaults set the model and effort a new conversation starts on.
- Anthropic’s suggestion: for roles where long open-ended work is most of the job, make Fable 5.1 the default “and consider setting an effort cap.” For mostly routine roles, keep Opus or Sonnet as the default and entitle Fable so users can switch when a task calls for it. Setup details are in the support articles linked from the video description; this wiki covers the admin side in Giving Admins More Visibility and Control Over Claude Spend.
How it compares with the creators’ rules.
- Same split, different starting point. Eliot, Robo Nuggets and Futurepedia start from Sonnet and move up. Anthropic starts from the kind of task and is willing to put Fable first for open-ended work. Both send routine work to cheaper models and hard open-ended work to Fable.
- Effort advice agrees. Anthropic’s “start lower and increase” matches Futurepedia’s medium default and Eliot’s warning about max. Note that Fable 5.1 defaults to High in Claude Code, so the “start at medium” advice means lowering it there.
- Brock’s game test (below) is a cost-per-task result. Sonnet 5.5 cost 5.7 times Sonnet 5 at the same token price. Anthropic’s framing predicts that per-token price alone would not show this.
- It predates Opus 5.5. Opus 5.5 now claims Fable-5.1-level work on most tasks at 20, which moves part of the “open-ended” row down to Opus. Ben AI’s seven tests below point the same way.
Anthropic’s November playbook for the Fable 5.2 lineup (added 2026-11-11)
Source: “Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku” (claude.dev, Addy Osmani, published 2026-11-09, ai-research/claude-dev-blog-which-claude-model-should-you-build-with-2026-11-11.md), published with Fable 5.2. Two creators summarised it: Rob at Switch to Dimension (raw/How_to_Prompt_Claude_Fable_5.2_Anthropic_s_Own_Guide.md, who names the author “Ari Mani”; the page credits Addy Osmani) and Ben AI (raw/Anthropic_Just_Revealed_How_to_Prompt_Fable_5.2.md).
It answers this article’s open question about whether Anthropic would still send open-ended work to Fable first now that Opus 5.5 exists: no, start on Opus 5.5.
| Task | Starting model | Starting effort |
|---|---|---|
| Everyday engineering and debugging, multi-PR features | Opus 5.5 | medium, then high |
| Classification, routing, tagging, high-volume extraction | Haiku 5.5 | medium |
| Subagents that search files, read logs or summarise | Haiku 5.5 or Sonnet 5.5 | Model default |
| Chat and interactive sessions | Sonnet 5.5 (Haiku 5.5 for simple chat at volume) | low or medium |
| Well-scoped coding: bug fixes, spec-driven features | Sonnet 5.5 | medium, then high |
| Everyday documents, slides, spreadsheets | Sonnet 5.5 (Opus 5.5 for a complex financial model or a large pile of sources) | high (medium in Claude Code) |
| Repeated agent tasks: investigation, review, drafting | Sonnet 5.5 | medium |
| Computer use and vision-heavy workflows | Opus 5.5 | medium |
| Major refactors and migrations, architecture planning, leading many subagents | Fable 5.2 | medium, then high |
| Deep research, scientific or quantitative analysis | Fable 5.2 | medium, then high |
| Not sure yet | Opus 5.5 | medium |
- Change effort before model. “Past
high, one more effort level can cost as much as a bigger model.” When a model falls short: (1) make sure it can verify its own work, (2) raise effort one level, (3) move up a model at a lower effort. - A higher price per token can mean a lower cost per task. On FrontierCode 2.0, Fable 5.2 at
highmatched Opus 5.5 atmax(45.3% vs 45.5%) for about half the cost per task, so test Fable before raising Opus pasthigh. - Re-run failures, not everything. On a SWE-bench Pro subset, Opus 5.5 at
lowwith failures re-run athighpassed about 97% for about 0.29 with every task athigh. - Switch at a natural break. “The prompt cache belongs to the model that wrote it,” and a move down a tier loses the earlier model’s thinking. This is the first-party version of Robo Nuggets’ rule below.
- Subagents on small models. Set
model: haikuormodel: sonnet; the built-in Explore subagent now runs on the session model, so define your ownExplorewithmodel: haiku. - Quick test in Claude Code: run one backlog task on two models with
/modeland compare/usage. - Subscriptions. Fable 5.2, 5.1 and 5 follow the same plan terms (Help Center,
ai-research/anthropic-support-claude-fable-models-on-your-plan-2026-11-11.md): included up to 50% of weekly limits on Max, Ultra and premium Team/Enterprise seats; usage credits on Pro and standard seats; not on Free.
Anthropic’s four-model video: one shop, four jobs (added 2026-11-11)
Source: raw/Fable_5.2_or_Opus_When_to_use_each_model_in_Claude_Code.md (“Fable 5.2 or Opus: When to use each model in Claude Code”, Anthropic, https://www.youtube.com/watch?v=C33hJGt-ySg, YouTube). It restates the November playbook with one worked example.
- Opus 5.5 at
medium: adding monthly subscriptions to an online shop (data model, API and UI). If it falls short, raise effort first: “a smaller step than changing the model, and it usually keeps the prompt cache.” - Fable 5.2: migrating the whole shop to a new architecture. Give it the goal and the non-negotiables (“Checkout can’t go down and orders can’t be lost”).
- Sonnet: a bug with a clear reproduction, where “I can say what done looks like in one sentence.”
- Haiku: sorting every incoming customer message through the Claude API.
- Mixed: Fable leads a framework migration and hands each package upgrade to Sonnet subagents.
- Brock Mesarich’s ^[inferred] follow-up test of Fable 5.2 against Opus 5.5 (
raw/I_Tested_Claude_Fable_5.2_vs_Opus_5.5_WILD_RESULTS.md, https://www.youtube.com/watch?v=AQWgE5LWQ4A) and three other launch-week comparisons are in Fable 5.2 vs Opus 5.5 head-to-head tests.
A second creator’s routing rules (Robo Nuggets, 2026-10)
Jay at Robo Nuggets lists routing as two of his 23 Claude 5.5 mistakes (raw/Claude_5.5_Model_Mistakes_you_may_still_be_making_23_Tips_and_Fixes.md, published 2026-10-12; full list in Claude 5.5 Mistakes and Fixes). They agree with Eliot’s levels and add two points:
- Sonnet for defined work, Opus for exploratory work. That is close to Eliot’s employee and specialist levels.
- Don’t switch models mid-chat to save money. Each model has its own prompt cache, so the switch re-reads the whole chat at full price. Jay relays an Anthropic engineer: 100,000 tokens into an Opus chat, switching to Haiku for an easy question costs more than letting Opus answer. Pick the model when you start the task.
- Let a router decide. Jay sends every prompt through Jev. Over three weeks it sent 9 in 10 of his prompts to Sonnet or Haiku, and the routing cost $0.13 in total. At list price Haiku 5.5 is about 40 times cheaper per token than Opus 5.5.
- Big models go big on open-ended prompts. Jay relays one user whose Fable session spun up 10 side agents and used his whole 5-hour limit in 5 minutes. That is a single report, but it is the same failure as Eliot’s fiancee’s Pro plan, at a higher tier.
A third creator’s rules (Futurepedia, 2026-08)
Source: raw/Use_Claude_Better_Than_99%_of_People_Beginner_to_Pro.md (Futurepedia, YouTube, https://www.youtube.com/watch?v=5aUgDDWd4-k, published 2026-08-16, Opus 5 era). Series summary in The Claude Ecosystem in Four Layers (Futurepedia).
- Haiku “basically never.” Sonnet is the daily driver on Pro; switch to Opus for harder reasoning. On Max he stays on Opus most of the time, especially in Cowork and Code, and prefers Sonnet for some creative tasks.
- Fable only when Opus is stuck. Opus 5 was strong enough that Fable was “way more token intensive for not that much extra.”
- Effort ladder: medium for most daily tasks, high for coding and analysis, extra high for agentic or multi-step coding sessions. Too much effort on a simple task takes longer and “might even be worse,” because it overcomplicates the answer.
- His medium default for daily work agrees with the Sonnet 5.5 and Fable 5.1 app defaults recorded above, not with Eliot’s “stick in high.”
A head-to-head build: Sonnet 5, Sonnet 5.5, Opus 5.5, Fable 5.1 (Brock Mesarich, 2026-09)
Source: raw/I_Tested_Sonnet_5.5_vs_Opus_5.5_WILD_RESULTS.md (Brock Mesarich, YouTube, https://youtu.be/pn08Kdp998Y, 8:55, published 2026-09-29).
Brock gave four models the same prompt: build a Minecraft-style 3D world from a list of hard constraints, planning the architecture before writing code. All four ran at high effort. Costs and times are his own measurements from one run each.
| Model | Cost | Time | His verdict |
|---|---|---|---|
| Sonnet 5 | $1.59 | 7 min 6 s | Weakest: glitchy, could walk under the lake, no clear goal |
| Sonnet 5.5 | $9.02 | 36 min 43 s | Best overall: smooth movement, rabbits and fish, weather and nightfall |
| Opus 5.5 | $11.95 | 41 min 47 s | ”Pretty solid”: consistent world, minor glitch in the water |
| Fable 5.1 | $17.45 | 40 min 30 s | Strong visuals, but the screen shook when walking |
- For a build with a detailed spec, Sonnet 5.5 beat both larger models in his eyes and cost about three-quarters of Opus 5.5 and half of Fable 5.1. That fits the employee level above for well-defined work.
- The new Sonnet is not cheap per task. It cost 5.7 times as much as Sonnet 5 at the same per-token price, because it did far more work. Budget by task, not by token price.
- Fable 5.1 was not slower than Opus 5.5 on this task (40.5 vs 41.8 minutes), but it cost about $5.50 more.
- Limits: one prompt, one run per model, a subjective verdict, and a game rather than office work. Brock says as much: “not everybody’s going to be generating video games.” For a method to run your own comparison, see Picking the Right Model.
Seven office tasks: Opus 5.5 vs Fable 5.1 (Ben AI, 2026-09)
Source: raw/I_Tested_Opus_5.5_vs_Fable_5.1_on_7_Real_Use_Cases_Not_Even_Close.md (Ben AI, YouTube, https://www.youtube.com/watch?v=3ogITvjOh30, 17:12, uploaded 2026-09-24). Added 2026-10-28.
Ben gave both models the same prompt for seven of his own business tasks, one run each. Times, token counts and costs are his own readings; he did not state the effort level or plan.
| Task | Fable 5.1 | Opus 5.5 | His verdict |
|---|---|---|---|
| Slide deck: last 30 days of marketing performance from his connectors, styled from reference images | 21 min, 20M tokens | 22 min, 30M tokens, “half the cost” | Opus: charts rendered properly, copy shorter and clearer; Fable’s charts and tables were broken or tiny |
| Landing-page redesign from a reference site, a liquid-glass GitHub repo and scroll animations | 14 min, 5M tokens | 17 min, 13M tokens, “almost half the price” | Opus: closer to the reference, smoother animations, added a comparison table and FAQ |
| Competitor research on YouTube titles and thumbnails from the last 7 days | 14 min, 50 videos analysed | 15 min, 168 videos (59 non-English), over twice the tokens, “slightly more” cost | Opus, for depth |
| Strategies for AI adoption from Claude’s customer case studies | 20 min, 13 strategies, $15 | 13 min, 15 strategies in four groups, about half the cost; it ran most of the reading on Sonnet subagents | Similar quality; Opus on cost and speed |
| Turn a YouTube video into a written lesson doc with annotated screenshots | 19 min, 9M tokens | 14 min, 13M tokens, $5, half the cost | Opus: better formatting and clearer copy |
| Customer-intelligence report from two months of support calls, Q&As and tickets | 1 h 15 min, 555 calls and 266 tickets, 33M tokens, every idea backed by a quote | 40 min, 248 calls, 7M tokens, links to transcripts but no quotes | Fable: deeper and better ideas. He gives the costs as “62” without saying which is which ^[ambiguous] |
| Business-trajectory review from his second brain | 17 min, about 200 documents read, $17 | 5 min, about 50 documents read, $2 | Fable: longer and more complete, with a “what moved the needle” section Opus left out |
- His routing rule: Opus 5.5 for “almost all use cases,” especially copy and design. Fable 5.1 only when a task has to read a lot of data or context and you accept the extra cost, because in his tests it spawned more subagents and read further.
- Fits the list prices. Opus 5.5 is 20 against Fable 5.1’s 50, so Opus can use more tokens and still cost about half, as it did on four tasks.
- A rumour he repeats. He passes on a theory that Anthropic runs new models at higher effort for the first weeks and then turns it down, and says Fable 5.1 seemed to go less deep than at launch. Neither he nor any source here gives evidence; Anthropic has published nothing of the kind. Treat it as unverified.
- Limits: one run per task, his own judgement, and costs read off a usage screen without the effort setting. Use the table to pick which of your own tasks to test, not as a ranking.
Try It
- Check the picker before you start. Open the model selector and confirm it shows Sonnet. Choose a higher model only when the task needs it.
- Sort this week’s tasks into the four levels. List ten recurring tasks and mark each as intern, employee, specialist or consulting team. Anything you have been running on Opus or Fable that is really an “employee” task is wasted allowance.
- Give Fable a goal, not instructions. Pick one open business question (“where am I undercharging?”) and give Fable only the outcome, with your connectors enabled. Compare the result with what Sonnet produces from the same prompt.
- Run the planner/executor split once. Have Opus 5.5 write a step-by-step plan, then start a fresh Sonnet 5.5 chat to carry it out. A fresh chat is cleaner than switching models mid-thread, which rereads the whole history without the cache (see Two-Model Workflow).
- Leave effort at the default for a week. Lower it only for quick lookups. Raise it only when the default fails on the same task twice.
- Price one open-ended task both ways. Following Anthropic’s cost-per-task rule, run the same research or analysis task on Opus 5.5 and on Fable 5.1 at medium, and compare total cost or share of your limit, not the per-token price.
- If you run a Claude org, set defaults by role. List your roles as mostly open-ended or mostly routine. Give the open-ended roles a Fable 5.1 default with an effort cap; give the routine roles an Opus or Sonnet default with Fable entitled.
Open Questions
- Haiku 5.5 in the apps. Haiku 5.5 shipped on the API on 2026-10-07 at 0.50 per Mtok for prompts up to 100K tokens and 2.50 above (Claude Haiku 5.5). The wiki has no source yet on whether the claude.ai and Cowork model picker switched to it. (Corrected 2026-10-27 after rebasing onto the W41 sweep; this line previously said Haiku 5.5 was unreleased and that the wiki had no Haiku article.)
- Fable 5.1 on subscriptions. Resolved 2026-11-11: the Help Center says Fable 5.2, 5.1 and 5 all follow the same terms (Max, Ultra and premium seats included up to 50% of weekly limits; Pro and standard seats on usage credits). See the November section.
- App defaults. Eliot calls High the default effort in the July 2026 app, while the wiki records Medium for the current Sonnet and Fable. It is unclear whether the default changed between releases or he misread the picker.
- Which plans have the three admin controls. Anthropic’s video does not say whether model entitlements, effort caps and defaults are Enterprise-only or also on Team. The support articles it links were not fetched for this article.
- Anthropic’s rule after Opus 5.5. Resolved 2026-11-11: the November playbook and the Fable 5.2 launch page say to start most work on Opus 5.5 and move up to Fable 5.2 only for the hardest technical work.
Related
- Which Effort Level to Use — the effort side of the same decision, per level
- The Claude Ecosystem in Four Layers (Futurepedia) — Futurepedia’s August 2026 model and effort picks in context
- Claude Sonnet 5.5 — the everyday-employee default and its near-Opus benchmarks.
- Claude Opus 5.5 — the specialist level; always-on thinking and Fable-class safeguards.
- Claude Fable 5.2 + Mythos 5.2 — the consulting-team level since 2026-11-09.
- Claude Fable 5.1 + Mythos 5.1 — the consulting-team level before that.
- Model vs. Effort — what the effort setting changes, in Claude Code.
- Picking the Right Model — the eval-based method once routing by intuition stops working.
- Two-Model Workflow — mechanics and limits of the planner/executor split.
- Claude Usage Limits Playbook — the companion guide to stretching a plan.
- Claude 5.5 Mistakes and Fixes — the habits checklist that includes the two routing rules above.
- Onboarding Module 1 — the short version of this routing for new users.
- Using Opus 5.5 as Your Daily Driver — Anthropic’s Claude Code version: medium by default, high for cross-file changes, Sonnet subagents.
- Giving Admins More Visibility and Control Over Claude Spend — the admin model-default and entitlement controls.