Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. In his launch-day video, Matt Wolfe reads the benchmark claims, runs three creative tests and tours demonstrations from early-access users. He had only a few hours with the model, so this is a first look, not a long-term reliability review. Here is what the official material establishes, what Matt actually tried and where the impressive demos still need scrutiny.
For routine coding and knowledge work, start with Opus 5.5, as Anthropic itself recommends. Try Fable 5.2 on a hard, open-ended technical problem where better planning or deeper investigation might justify extra latency and spend. Mythos 5.2 is a restricted variant for verified cyber and life-science work, not a general upgrade button. Neither a vendor benchmark nor a viral one-prompt demo proves a production workflow will succeed.
Watch Matt Wolfe's First Look
Matt Wolfe's video and the supplied transcript are the source for his test conditions and opinions. Product specifications below are checked against Anthropic's 9 November launch post. Pricing and availability can change; checked 10 November 2026.
Fable and Mythos Share a Model, Not an Access Policy
Anthropic says Fable and Mythos 5.2 use the same underlying model. Fable carries stronger cybersecurity and biology safeguards and is available broadly. Mythos is limited to organizations accepted into its Cyber or Life Sciences Verification Programs. Those restrictions matter when reading a chart: some cyber and protein-design scores are specifically for Mythos, not the public Fable experience.
The company's own positioning is narrower than the hype: Opus 5.5 is its recommended daily driver. Fable is meant for ambitious, open-ended technical work and can act as an adviser for another Claude model or orchestrate smaller agents. The launch also claims up to 30% faster output than Fable 5.1; that is a vendor comparison, not a guarantee for every prompt.
The Benchmark Improvement Is Real in the Chart, but Context Matters
Anthropic's published evaluation table reports the following results. These are Anthropic-reported, with different effort settings and benchmark conditions; they are not our reproduced measurements.
| Evaluation | Fable 5.2 | Opus 5.5 | How to read it |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% at Xhigh | 45.5% | Agentic code-change test. Compare effort and cost per successful task, not score alone. |
| FrontierSWE v2 | 69.9% | 62.3% | A stronger result on this software-engineering evaluation, not a guarantee on your repository. |
| Terminal-Bench-Science 0.1 | 74.6% | 58.7% | Large gain on the published scientific-agent tasks; Mythos scores 75.3% separately. |
| AA-Briefcase v1.1 | 66.1% rubric; 1785 Elo | 61.2% rubric; 1800 Elo | Even the vendor table is mixed: Fable leads on rubric, Opus on Elo. |
Anthropic also says Fable at High effort scored 45.3% on FrontierCode 2.0, approximately matching Opus at Max effort (45.5%) for roughly half the cost per task in that test. This is a useful efficiency hypothesis, not a blanket statement that Fable is cheaper. In a separate assessment, Artificial Analysis recorded a top Intelligence Index position for its Max/default-fallback configuration but also measured a slow, costly run: about $26.54 and 46.4 minutes per Intelligence Index task at the snapshot checked 10 November. That composite task cost is not the API price of a normal user request, and its rank may move.
Access, API Pricing and a Sensible First Test
Anthropic lists standard API pricing at $8 per million input tokens and $40 per million output tokens. Optional fast mode is $16/$80 and is advertised as producing tokens up to 2.5 times faster. The API model ID is claude-fable-5-2; Anthropic also lists partner-cloud access. Eligible paid plans have a launch-week usage promotion, but plan limits differ, so check your own model picker and usage page rather than treating an API rate as your subscription bill.
For a first comparison, take one genuinely difficult task with a pass/fail check, run it under the same tools and time budget with Opus and Fable, and log total tokens, wall time, revisions and final correctness. Do not give either model unrestricted production credentials for a launch-day experiment. If Fable only makes a prettier prototype while requiring more review, the added capacity has not yet paid for itself.
Matt Wolfe's Three Hands-On Tests
BuseyBench: Matt asked for an SVG likeness of Gary Busey. Fable spent about 19 minutes 50 seconds, used roughly 122,000 tokens and cost close to $5 in his run. He did not rank the result first by his own judge. This is one creative-output test with a particular prompt and judge, not a general vision benchmark.
Game prototype: His Megabonk-inspired request produced an original game he called Shardstorm in roughly 16 minutes at Medium effort. He liked the mechanics but thought the visual finish lagged an earlier Opus 5.5 example. The model also steered away from copying protected game assets. That is a prototype to inspect and iterate, not proof of a shippable game.
Gadget concept: Matt asked for a pixel-character desk companion. Fable returned a concept, parts list, Pico-based firmware, case files and assembly instructions. In the video, the device was not built or physically validated. This matters because Anthropic's own electronics case study used human engineers and manufacturing partners to assemble three working prototypes. The two examples are not equivalent evidence.
What the Bigger Demonstrations Actually Show
Anthropic reports an internal operating-system prototype built over eight days with more than 1,000 subagents, and a collaboration with Diode Computers that produced working speakerphone, voice-hub and virtual-pet prototypes. The hardware article explicitly says the devices are early prototypes with unknown long-term durability or commercial certification. These are unusually ambitious, company-led demonstrations with supporting people, tools and time, not a one-click feature any subscriber can assume.
Matt also tours early-access community creations. They are valuable for seeing the range of prompts, but their authors' posts do not establish controlled model comparisons. The linked demo is evidence of an output, not of the complete prompt, edit history, cost or reliability.
| Creator | Example | What to check |
|---|---|---|
| Matthew Berman | Gears-inspired game footage | Playability, original assets and how much hand iteration followed the prompt. |
| Alex | Montage of games, animation and creative-app experiments | Which examples are runnable and which are short visual showcases. The supplied status link was truncated. |
| Peter Yang | Code-generated Fushimi Inari travel animation | Animation quality and the underlying project, beyond a 20-second clip. |
| Future Brian | Dragon-riding simulation, plus game and video-editing experiments | The author also reports failures on some other experiments, useful counterweight to the highlights. |
| Matt Shumer | Mirror Earth game demo | Try the actual build. Matt Wolfe observed glitches during his brief test. |
Which Model Should You Try First?
| Your job | Start with | Reason |
|---|---|---|
| Everyday coding, writing or research | Opus 5.5 | Anthropic recommends it as the daily driver; establish a baseline before paying for deeper reasoning. |
| A hard, open-ended technical investigation | Fable 5.2, alongside an Opus baseline | This is the release's target use case. Judge the verified result, latency and all-in cost. |
| Cybersecurity or biomedical research needing Mythos | Apply through Anthropic's verification route | Mythos access is restricted; do not assume Fable's safeguards behave identically. |
Sources and Demo Links
Anthropic launch and benchmark methodology · Anthropic/Diode hardware case study · Artificial Analysis model page · Matt Wolfe's full video. Community posts are linked in the table above. Source checks were made on 10 November 2026; model access, pricing and live rankings should be rechecked before purchase or deployment.