Claude Fable 5.1 for Cursor and Claude Code: Same $10/$50 Price, 75% Cheaper Cache Reads — and a Cline Default Trap

By AICoderScope Team · 14 min read
claudeanthropiccursorclineclaude-codepricingcost-analysisapi

TL;DR: Claude Fable 5.1 (released September 1, 2026) costs the same $10/$50 per million tokens as Fable 5, but cache reads dropped from $1.00 to $0.25 per million — which cuts a long cached agentic session by roughly a third. If you were already paying for Fable 5, switching is free money. If you weren’t, Opus 5 at half the price is still the right default for most coding.

Claude Fable 5.1Claude Opus 5Claude Sonnet 5
Best forMulti-hour agentic refactors, long-horizon researchComplex daily coding — the Claude Code default on Max/APIRoutine coding on a per-token budget
Input / output per MTok$10 / $50$5 / $25$2 / $10
Cache read per MTok$0.25 (0.025x)$0.50 (0.1x)$0.20 (0.1x)
Heavy cached session (math below)~$8.00~$6.25~$2.50
The catchSlowest in the lineup; no temperature control; breaking API changes vs Fable 5~Half the Terminal-Bench 4.0 headroom on marathon tasksWidest quality gap on the hardest 20% of work

Honest take: Fable 5.1 is an upgrade you take if you were already on Fable 5 — same sticker price, strictly cheaper in cached agentic loops, better at exactly the long-running work Fable is for. It is not a reason to move off Opus 5 or Sonnet 5. And if you run Cline with an Anthropic key, go check your model setting right now: since v4.1.17 the provider default resolves to Fable 5.1, at double Opus 5’s rates.

What is Claude Fable 5.1 and what changed from Fable 5?

Claude Fable 5.1 is Anthropic’s top “Mythos-class” model, released September 1, 2026 as the successor to Fable 5. The API model ID is claude-fable-5-1 — a dateless pinned snapshot, no date suffix. Specs verified against Anthropic’s model docs on September 16, 2026: 1M-token context window (default and maximum, billed at standard per-token rates across the whole window), 128K max output tokens, adaptive thinking always on, default effort high, reliable knowledge cutoff June 2026, and a retirement commitment of no sooner than September 1, 2027. Its restricted sibling, Claude Mythos 5.1, has identical specs and pricing but is invitation-only through Anthropic’s Project Glasswing.

Three things actually changed for coding use:

  1. Cache reads cost 0.025x base input instead of 0.1x — $0.25 per million tokens instead of Fable 5’s $1.00. Cache writes ($12.50/M for 5-minute, $20/M for 1-hour) and the 512-token minimum cacheable prompt are unchanged.
  2. The gains concentrate in long-horizon agentic work. Anthropic’s own numbers (vendor-run, not independently verified): Terminal-Bench 4.0 goes from 42.0% to 55.8%, Terminal-Bench-Science from 24.7% to 52.6%, AutomationBench from 17.1% to 31.4%. Multilingual performance is flat versus Fable 5.
  3. Three breaking API changes if you call Fable 5 directly — covered in the migration section below, because two of them can bite Cursor/Cline BYOK users with custom harnesses.

How much does Claude Fable 5.1 cost per session?

Same as Fable 5 on every line except cache reads. All prices verified against Anthropic’s official pricing page on September 16, 2026:

Per million tokensFable 5.1Fable 5Opus 5Sonnet 5
Input$10.00$10.00$5.00$2.00
Output$50.00$50.00$25.00$10.00
Cache read$0.25$1.00$0.50$0.20
5m cache write$12.50$12.50$6.25$2.50
Batch input / output$5 / $25$5 / $25$2.50 / $12.50$1 / $5

For an uncached 200K-input / 25K-output session — the shape we used in the Opus 5 default cost analysis — Fable 5.1 runs $3.25 ($2.00 input + $1.25 output), identical to Fable 5, versus $1.63 on Opus 5 and $0.65 on Sonnet 5.

The cache cut only shows up in sessions that re-read a cached prefix many times — which is exactly what Claude Code, Cursor agents, and Cline loops do. Take a multi-hour agentic session that accumulates 6M tokens of cache reads, 200K of 5-minute cache writes, 100K of fresh input, and 60K of output:

Cost componentFable 5Fable 5.1Opus 5Sonnet 5
6M cache reads$6.00$1.50$3.00$1.20
200K cache writes$2.50$2.50$1.25$0.50
100K fresh input$1.00$1.00$0.50$0.20
60K output$3.00$3.00$1.50$0.60
Session total$12.50$8.00$6.25$2.50

That’s 36% off versus Fable 5 for this shape — consistent with Anthropic’s launch estimate of roughly 25% savings on typical workloads and up to about 45% on heavily agentic ones (Anthropic’s numbers, as reported at launch; your ratio of reads to output decides where you land). At two such sessions per weekday, the monthly bill drops from about $550 on Fable 5 to about $352 on Fable 5.1.

One number in that table deserves a second look: Fable 5.1’s cache read ($0.25/M) is now half of Opus 5’s ($0.50/M), even though its base input price is double. In read-dominated marathon sessions, the Fable-over-Opus premium shrinks from 2x to about 1.3x ($8.00 vs $6.25 above). The premium never inverts — writes, fresh input, and output still cost 2x — but the longer and more cache-heavy the session, the cheaper Fable 5.1 gets relative to everything else. Anthropic priced this model to be run for hours.

Is Fable 5.1 actually better at coding?

On long agentic tasks, yes, by a wide margin — with the caveat that every number below except CursorBench is Anthropic-run:

BenchmarkFable 5Fable 5.1
Terminal-Bench 4.0 (agentic terminal coding)42.0%55.8%
Terminal-Bench-Science 0.124.7%52.6%
AutomationBench (multi-step office tasks)17.1%31.4%
CursorBench 3.2 (Cursor-run, max effort)—73.4%

Cursor’s team, who ran their own eval rather than quoting Anthropic’s, called it “the most capable model we’ve run on CursorBench 3.2” and specifically credited its ability to verify its own work across a difficult task from start to finish. That matches Anthropic’s framing: the improvement is concentrated in sessions that run long and use tools heavily. On short, well-scoped tasks — the kind Sonnet 5 already handles — the published deltas are small, and no independent SWE-bench Verified score for Fable 5.1 existed as of September 16, 2026. If your work is mostly single-file edits, the benchmark story gives you no reason to pay 5x Sonnet’s rate.

How do you use Fable 5.1 in Cursor?

Claude Fable 5.1 is in Cursor’s native model picker — no BYOK required. Cursor announced availability at launch and maintains a dedicated docs page for the model (cursor.com was unreachable from our verification environment this run; availability confirmed via Cursor’s own announcement and its docs index, both dated early September 2026). For BYOK through an Anthropic API key, the model string is claude-fable-5-1.

Two practical cautions for Cursor users:

  • Cursor ran CursorBench at max effort. Fable 5.1’s default effort is high and adaptive thinking is always on and billed as output tokens. The 73.4% headline was achieved at the model’s most expensive setting; expect the everyday experience at default settings to sit somewhat below it.
  • It’s the slowest model in Anthropic’s lineup (“slower” comparative latency in the official model table). For inline, interactive edits, that latency is the wrong trade — this is a background-agent and long-task model.

How do you use Fable 5.1 in Claude Code?

It’s available but never the default — on any plan or provider. Verified against the Claude Code model-config docs on September 16, 2026: Max, Team Premium, Enterprise, and pay-as-you-go API sessions default to Opus 5; Pro and Team Standard default to Sonnet 5 (full breakdown in our Opus 5 default analysis). To use Fable 5.1 you select it explicitly:

# one session
claude --model fable

# inside a session
/model fable

# persistent
export ANTHROPIC_DEFAULT_MODEL=fable

Requirements and gotchas, all from the official docs:

  • Claude Code v2.1.257 or later resolves the fable alias to Fable 5.1 (v2.1.219–2.1.256 resolve it to Fable 5). Check with claude --version.
  • The best alias picks Fable when your plan can reach it, otherwise Opus — useful in shared configs where some seats have Fable access and some don’t.
  • On Max plans and premium Team/Enterprise seats, Fable-family usage is capped at 50% of your weekly usage limits; past that you either buy usage credits or drop to another Claude model (per Anthropic’s “Claude Fable models on your plan” support page, checked September 16, 2026). Fable 5.1 burns that allowance at the same rate Fable 5 did — the cache discount helps your API bill, not your subscription cap. We covered the credits mechanics in the Fable 5 credits cost breakdown.

The Cline trap: your default model may have silently become Fable 5.1

This is the one item in this article that can cost you money by doing nothing. Cline’s v4.1.17 release (September 2, 2026, per the official GitHub releases page) shipped a catalog refresh across 57 providers, and its release notes state the Anthropic provider “now resolves to Claude Fable 5.1 instead of Claude Opus 5.”

Translation: if you use Cline with an Anthropic API key and never pinned a specific model, your agentic loops may now run on a $10/$50 model instead of a $5/$25 one — and Cline sessions are long, tool-heavy, and output-heavy, the exact shape where the absolute dollar difference is largest. The cache-read discount softens this (see the session table above: $8.00 vs $6.25 for a heavy cached session), but softer is not free, and on uncached or write-heavy work the gap is a clean 2x.

The fix takes thirty seconds: open Cline’s settings, find the Anthropic provider’s model dropdown, and pin claude-opus-5 or claude-sonnet-5 explicitly. Pin — don’t rely on the default, which has now changed twice in 2026. If you actually want Fable 5.1 for a marathon refactor, select it deliberately and switch back after. Our Cline review covers the provider settings layout.

What breaks if you migrate from Fable 5?

Three API-level breaking changes, verified from Anthropic’s what’s-new page. These matter if you call the API directly or run a custom harness; Claude Code, claude.ai, and the Agent SDK handle them for you.

  1. Forced tool use returns a 400 error. tool_choice: {"type": "any"} and {"type": "tool", ...} are rejected. Thinking is always on, and a forced call would skip it. Use strict: true tool schemas or structured outputs instead. If your Cline/Cursor custom mode or MCP wrapper forces tool choice, it will hard-fail on Fable 5.1.
  2. Thinking blocks are one-directional. Fable 5.1 can read earlier models’ thinking blocks; no earlier model can read Fable 5.1’s. A router that falls back mid-conversation from Fable 5.1 to Opus 5 silently drops the reasoning (unbilled, but gone).
  3. Editing earlier turns invalidates thinking blocks. Rebuilding the system prompt, tools array, or any earlier message mid-conversation gets requests rejected with a 400 (enforced for accounts created on or after August 31, 2026). Harnesses that inject-and-remove per-turn reminders into history break; the sanctioned replacement is turn-scoped system messages (beta).

Two behavior changes also affect cost without breaking anything: Fable 5.1 batches parallel tool calls less consistently than Fable 5 (more round trips in agent loops unless you add an explicit batching instruction to your prompt), and it’s more prone to rewriting a whole file where a targeted edit would do — which costs output tokens at $50/M. Both have documented prompting fixes in Anthropic’s Fable 5.1 prompting guide. Also unchanged and still surprising to API newcomers: no temperature, top_p, or top_k (400 error), and no assistant prefill.

When should you NOT use Fable 5.1?

  • Routine daily coding. Sonnet 5 at $2/$10 handles it, and the benchmark gap concentrates in marathon tasks you’re not running. This is most people, most days.
  • Interactive, latency-sensitive editing. Slowest model in the lineup, thinking always on. Wrong tool for tab-completion-adjacent workflows.
  • Anything under a compliance requirement for zero data retention. Fable 5.1 carries mandatory 30-day retention and isn’t available under zero-data-retention terms unless Anthropic expressly authorizes it — a real blocker for some enterprise teams that can use Opus 5 fine.
  • Privacy-first or zero-marginal-cost setups. No local option exists at this tier; if the requirement is code-never-leaves-the-machine, a 24GB-GPU local stack is the answer — see runaihome’s best local models by VRAM guide and the Ollama review on aifoss.dev.
  • If you’re on Fable 5 with a hand-built harness and no time to test. The price cut is real, but the three breaking changes above mean “just change the model string” isn’t a safe Friday-afternoon deploy.

If you’re comparing across vendors rather than within Anthropic’s lineup, our AI code editor cost comparison and Sonnet 5 review have the wider table.

FAQ

Is Fable 5.1 more expensive than Fable 5? No. Base input, output, cache writes, and batch prices are identical; cache reads are 75% cheaper ($0.25/M vs $1.00/M). For any workload, Fable 5.1 costs the same or less than Fable 5 at list prices.

Is Fable 5.1 the default model anywhere? No. It is not the default in Claude Code on any plan or provider, and Cursor’s default is unchanged. The one default that did flip is Cline’s Anthropic provider (v4.1.17, September 2, 2026) — which is why this article exists.

Does the 1M context window cost extra? No. Anthropic bills the full 1M window at standard per-token pricing on Claude 4.6-and-later models, including Fable 5.1. There is no long-context surcharge tier.

What’s the difference between Fable 5.1 and Mythos 5.1? Same model capabilities, specs, and pricing. Fable 5.1 is generally available with Anthropic’s production safeguards; Mythos 5.1 is restricted to approved organizations in Anthropic’s Project Glasswing program.

Can I turn off thinking to save output tokens? No — adaptive thinking is always on and thinking: {"type": "disabled"} returns a 400 error. You can steer depth with the effort parameter (default high), and on Fable 5.1 you can now change effort mid-conversation without invalidating the prompt cache (beta header required).

Which older articles does this supersede? Pricing and defaults here update what we wrote in the Fable 5 coding review and the Fable 5 return coverage — those remain accurate for Fable 5 itself.

Sources

Last verified September 16, 2026. Anthropic docs (platform.claude.com, code.claude.com, support.claude.com) and Cline’s GitHub releases were fetched directly; cursor.com, venturebeat.com, and llm-stats.com were unreachable from our verification environment this run, so Cursor availability rests on Cursor’s own announcement and docs listing, cross-checked against independent coverage. Pricing and features change frequently; verify current state before committing to a model.

Was this article helpful?

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.