DeepSWEmini-swe-agent
70%±4%
- bench
- DeepSWE
- version
- v1.1
- split
- public 113 tasks
- harness
- mini-swe-agent
- effort
- max
- n
- 113
- metric
- Pass@1
- value
- 70%
- ci
- ±4%
- $/task
- $21.63
- tokens
- 119k out
- steps
- 88
claude-fable-5Mythos-class Claude. OpenRouter copy: built for autonomous knowledge work and coding; text, image, and file in; text out; 1M context. Continuum picker ids often prefix continuum/; the gateway id is the bare claude-fable-5.
In Claude Code, /model fable and the best alias resolve here when the org has access. Effort ladder is low, medium, high, xhigh, max. Default effort is high. Thinking cannot be turned off. Tool use, structured outputs, prompt caching, and Batch API ship with the first-party SKU.
Anthropic launched Fable 5 on 9 Jun 2026, disabled it on 12 Jun after a US export-control directive, and redeployed it on 1 Jul 2026. Source: Anthropic’s Fable 5 / Mythos 5 post, read 19 Aug 2026.
Use Fable when the job is larger than one sitting: a multi-file design that has to stay coherent overnight, or a bug that already beat Sonnet and Opus once. Do not use it as the model you leave selected all week.
DeepSWE official (mini-swe-agent, 113 tasks, Best / All effort levels table) puts Fable at 70% Pass@1 ±4% at max, $21.63 per task, 119k output tokens, 88 steps. Opus 5 is 74% at $11.84. Sol is 73% at $8.39. Fable is not the accuracy leader on this board, and it is the most expensive row we fetched.
vals.ai Fable 5 card, fetched 19 Aug 2026: Index 66.04% ±1.03, $28.80 per Index test, 37 min 51 s latency. Those are Index-page extras, not DeepSWE, and not hero chips. The per-bench bars on that HTML were placeholders (0.0%) in this fetch. We will not invent SWE-bench or Terminal-Bench percents from them.
Ranked view: Best coding models: the independent leaderboard puts this row and every other card that carries an independent coding score on one board, with the confidence intervals left visible.
Copied from AA or vals HTML we opened. 0.0% placeholder bars are omitted. Do not average these into the hero tiles, and do not invent CI, $/task, or steps for Artificial Analysis.
| Bench | Printed | Note | Source | URL | As of |
|---|---|---|---|---|---|
| GDPval-AA v2 | 61.9% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| τ³-Banking | 38.1% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| Terminal-Bench v2.1 | 84.6% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| SciCode | 60.2% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| Humanity's Last Exam | 55.5% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| GPQA Diamond | 92.6% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| CritPt | 28.6% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| AA-Omniscience Accuracy | 65.4% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
| AA-LCR | 76.7% | Printed on the AA Intelligence Evaluations grid (Artificial Analysis Claude Fable 5). Not a DeepSWE chip. | Artificial Analysis Claude Fable 5 | source | 2026-08-19 |
claude-sonnet-5). Opus 5 is the reasoning step before Fable ($5 / $25, 74% DeepSWE).Get Plus if you want Continuum to host Fable after the Claude weekly slice is gone. Not a Mac download.
Anthropic does not publish Fable 5 weights. There is no official Hugging Face repo and no Files button on this page. Get Plus buys hosted inference, not a local checkpoint.
anthropic/claude-fable-5 (Anthropic, Claude Platform on AWS, Amazon Bedrock, Azure BYOK, Google Vertex, Google Vertex Europe). We will not reprint their latency table.Lab card and OpenRouter listing only. No Hugging Face Files button, because there is no official repo to point at.
Fetched OpenRouter catalog on 2026-08-19. hugging_face_id was empty.
Continuum hosts claude-fable-5 on the live public allowlist fetched 19 Aug 2026. Plus and above. First-party Claude Code still bills your Anthropic plan.
We will not list fifteen OpenRouter providers. Six hosts exist on that page today; Continuum is a seventh path with a hosted id. If you want an aggregator, the slug is anthropic/claude-fable-5.
claude-fable-5
click to select
We list first-party and OpenRouter, plus Continuum only when the live public allowlist named the id. This is not a 15-host routing table.
DeepSWE official cost at the cited max effort is $21.63 per task. That is the hero economics number, because it includes the harness, the step count, and the output tokens (119k).
Official Anthropic pricing.md (canonical URL after redirect: platform.claude.com), fetched 19 Aug 2026: Fable 5 base input $10 / MTok, 5-minute cache writes $12.50 / MTok, 1-hour cache writes $20 / MTok, cache hits & refreshes $1 / MTok, output $50 / MTok. Batch API is half: $5 / $25 per million. Sonnet 5 on the same table is $2 / $10. Opus 5 is $5 / $25.
OpenRouter catalog and Continuum rate card, 19 Aug 2026, reprint the same list pair ($10 / $50) plus image input $1.00 /M and web search $0.01 per call on the catalog. Those extras are OR fields, not extra Anthropic columns.
OpenRouter’s paid-vs-list gap is the cache: weighted average input $2.807 /M against a $10 list, output essentially $50. Do not treat $10 as what a cached agent run pays.
On a Claude plan? The plan picker models the weekly caps.
Artificial Analysis cost/speed footnote, fetched 19 Aug 2026 from their Fable 5 page: list $10 / $50, about $3.14 per AA task and 71.1 tok/s. AA’s own title is “Claude Fable 5 (with fallback)”; the meta description names Adaptive Reasoning, Max Effort, Opus 4.8 Fallback. That is AA’s evaluation config, not a Continuum hosted alias. Intelligence Index v4.1.1 is nine evals. The integer 62 is the tile. The nine printed Index percents from live currentModel sit in the named-benches table, not as extra hero chips. Do not average AA cost with DeepSWE.
| Claim | Source | As of |
|---|---|---|
| DeepSWE 70% ±4% at max | DeepSWE official board | 2026-08-13 |
| Official Fable 5 $10 / $12.50 5m write / $20 1h write / $1 hits / $50 out; batch $5 / $25 | Anthropic pricing.md (platform.claude.com) | 2026-08-19 |
| OR catalog reprints $10 / $50 plus image $1.00/M and web search $0.01 | OpenRouter Fable 5 page + catalog | 2026-08-19 |
| Six OR hosts; P50 3.46s / 65 tok/s; weighted paid $2.807 / $49.97 | OpenRouter Fable 5 Providers + Pricing + Performance | 2026-08-19 |
| AA title “Claude Fable 5 (with fallback)”; meta names Opus 4.8 Fallback; Index 62; nine currentModel Index percents in named-benches | Artificial Analysis Claude Fable 5 | 2026-08-19 |
| vals Index 66.04% ±1.03; $28.80/test; 37 min 51 s | vals.ai Claude Fable 5 | 2026-08-19 |
| Fable is 50% of weekly pool on Max/premium | Anthropic: Claude Fable 5 on your plan | 2026-08-19 |
| Launch 9 Jun 2026; disabled 12 Jun; redeployed 1 Jul | Anthropic: Claude Fable 5 and Claude Mythos 5 | 2026-08-19 |
| Hosted id claude-fable-5 | GET /v1/chat/hosted/models/public | 2026-08-19 |
| OpenRouter id anthropic/claude-fable-5, context 1,000,000 | OpenRouter /api/v1/models | 2026-08-19 |
Same official DeepSWE harness (mini-swe-agent, public 113 tasks), fetched 2026-08-19. Different effort labels are the lab's own setting on that board, shown here rather than normalized.
Continue through Anthropic's model family with Claude Opus 5.
The official board is one harness and one split. Fable is built for long autonomous runs, not for winning a 113-task software-engineering set at any price. On this board Opus is both cheaper per task and slightly ahead. That is why the when-not block exists.
No. Anthropic serves it. Continuum hosts the same id for Plus and above. Claude Code on your own Anthropic login is still first-party.
Lab posts pick a harness, an effort, a split, and sometimes a private eval set. DeepSWE publishes the official mini-swe-agent row with cost, tokens, and steps. Scale SWE-bench Pro only counts when the public shared-harness board has a row we can fetch. A higher lab number is usually a different test, not a better one.
Verified against the live public allowlist on 2026-08-19. Base URL is https://continuumcode.ai/v1. Keys are cont_sk_ from Settings, Account, Inference API. Personal keys need Plus or above.
curl https://continuumcode.ai/v1/chat/completions \
-H "Authorization: Bearer $CONTINUUM_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"claude-fable-5","messages":[{"role":"user","content":"Review this diff."}]}'