[AINews] Anthropic Claude Fable 5 — 神話的だが安全、論争的な条件付き

By some measures, Opus 4.8, barely two weeks old, was already the leading model in the world. But now, 34 days after the SpaceXai deal and 63 days after the original Mythos announcement*, we have a Mythos-class model (at least 2x size of Opus) available to everyone (in coinciding with Claude Tokyo). It is a feat of incredible engineering (and commitment to access) to make these research models GA, and the benchmarks are great… with asterisks. Here they are on yesterday’s brand new, out of distribution, FrontierCode Diamond, going from 13.4% to 29.3%:

tweet

The blog and the system card contain most of the authoritative information, but don’t miss the youtube videos showing it playing Factorio, Pokemon (unlike Claude Plays Pokemon, this is just using vision, no complex harness as we covered in our pod), EDM visualization (never having head music before), 3D CAD editor creation and printing and more from their main intro video.

API pricing is also fantastic, at roughly 2x Opus.

The asterisks come because Fable is released with two controversial changes:

No ZDR: “We will require 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces. We won’t use this data to train new Claude models, or for any non-safety-related purpose, and we’ve instituted new privacy protections including logging all human access to the data and ensuring its deletion after 30 days in almost all cases ...” (see full policy)

RSI suppression: “In light of the ability of recent models to accelerate their own development, we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design). Using Claude to develop competing models already violates our Terms of Service, but enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms.

Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user. Fable 5 will not fall back to a different model. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT). These interventions will not affect the vast majority of coding work. We estimate they will impact ~0.03% of traffic, concentrated in fewer than 0.1% of organizations”.

The vast majority of users will not be affected by these limitations, but the open AI community is understandably upset, as you will see below.

You can find more of their recommendations on usage in Diane Penn’s Tokyo talk, which we have clipped below.

*(and 1 week-1 day after both Anthropic and OpenAI filed their S-1’s ahead of SpaceX’s IPO next week…)

AI News for 6/8/2026-6/9/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Top Story: Anthropic Claude Fable 5 and Mythos 5 release

What happened

Anthropic released two versions of its next major model family: Claude Fable 5 for general availability and Claude Mythos 5 for restricted access.

Anthropic officially announced Claude Fable 5 as its “first generally available Mythos-class model,” saying it exceeds any model it has previously made broadly available and is state-of-the-art on nearly all tested benchmarks @claudeai, @claudeai

Anthropic said Fable 5 is the same underlying model as Mythos 5 with added safeguards, and that some cyber/bio/chemistry/distillation-related prompts may be routed to Claude Opus 4.8 instead @ClaudeDevs, @scaling01

Anthropic stated that for a “narrow range” of potentially harmful topics, queries transparently fall back to Opus 4.8, and claimed 95%+ of sessions never see one according to early user-facing messaging @claudeai, @mikeyk

Anthropic developer messaging said fallback is available server-side and via SDK middleware in Python, TypeScript, Go, Java, and C# @ClaudeDevs

Pricing for both Fable 5 and Mythos 5 was reported as $10 / million input tokens and $50 / million output tokens; cache pricing was later reported by third-party evaluators as $12.50 / million cache writes and $1 / million cache reads @scaling01, @ArtificialAnlys

Fable 5 kept Anthropic’s 1M-token context window according to Artificial Analysis @ArtificialAnlys

Anthropic put Fable 5 into Pro, Max, Team, and seat-based Enterprise plans until June 22, then said it would require usage credits due to capacity constraints, with plans to restore broader subscription access later @ClaudeDevs, @scaling01, @ArtificialAnlys, @kimmonismus

Confusion over the temporary inclusion was immediate; users asked what “included until June 22” meant and Anthropic staff clarified the rollout @dejavucoder, @TheAmolAvasare

Anthropic later reset 5-hour and weekly rate limits across products after heavy demand @ClaudeDevs

Official claims and third-party benchmark data

Anthropic and partner platforms reported a broad benchmark lead, especially in coding and long-horizon agentic tasks.

Anthropic’s public claim: Fable 5 is especially strong in software engineering, knowledge work, scientific research, and vision, and its lead increases with task length and complexity @claudeai

Cursor said Fable 5 set a new CursorBench SOTA at 72.9%, 8 points above the previous best @cursor_ai

Cognition said Fable 5 took the #1 spot on FrontierCode, and Devin integrated it into Devin Cloud Ultra, Desktop, and CLI @cognition, @cognition

Cline reported Fable 5 at 88.0% on Terminal-Bench 2.1, beating GPT-5.5 by 4.6 points @cline

Artificial Analysis placed Fable 5 #1 on its Intelligence Index at 64.9, roughly 5 points ahead of GPT-5.5, and said Anthropic occupied the top two spots @ArtificialAnlys

Artificial Analysis also reported:

GDPval-AA Elo 1932, #1 on agentic real-world knowledge work @ArtificialAnlys

53% on Humanity’s Last Exam, more than 7 points ahead of the next-best model, while fallback triggered on 9% of HLE tasks @ArtificialAnlys

~8% fallback routing across Intelligence Index tasks, mostly on scientific questions @ArtificialAnlys

Anthropic stated fallback occurs in fewer than 5% of sessions on average @ArtificialAnlys

Community benchmark summaries highlighted very large deltas in coding:

SWE-Bench Pro: Fable 5 80.3% vs GPT-5.5 58.6% @Yuchenj_UW

FrontierCode Diamond: Mythos 5 30.9% vs second-best 13.4% @scaling01

Anthropic ECI 161.29 for Mythos 5 @scaling01

Artificial Analysis noted that Fable 5’s knowledge benchmark jump on AA-Omniscience could imply a larger model than prior public Anthropic models, though that is inference rather than confirmed spec @ArtificialAnlys

Product behavior, usage profile, and deployment details

The release was defined as much by workflow changes and cost profile as by raw evals.

Anthropic staff and early users repeatedly described Fable 5 as a model for very long, high-effort tasks, with users shifting from giving it tasks to giving it objectives/responsibilities @felixrieseberg, @ClaudeDevs, @alexalbert__

Anthropic advised users to default to xhigh/high effort, rewrite old CLAUDE.md instructions, and let the model use more judgment @alexalbert__

Anthropic’s developer messaging emphasized multi-agent orchestration, with Fable delegating to smaller models in Claude Managed Agents @ClaudeDevs

Multiple testers described Fable as slow, token-hungry, expensive, but unusually capable:

Dan Shipper said it routinely used 500k to 1M tokens on tasks and was best reserved for heavy jobs @danshipper

Simon Willison called it “slow, expensive and capable” @simonw

Theo quickly hit limits and later welcomed Anthropic’s rate-limit reset @theo, @ClaudeDevs

Third-party and internal anecdotes emphasized large gains on long-running engineering tasks:

Ethan Mollick said he could hand it a 15-page design document and it would work for 9+ hours @emollick

Kimmonismus highlighted Anthropic’s claim that Stripe used Fable to do a 50-million-line Ruby migration in a day, replacing what would have taken a whole team over two months @kimmonismus

Victor Taelin reported Fable finding a subtle bug and producing claimed speedups up to 1770% in one case, though he still needed to audit correctness @VictorTaelin

Anthropic-associated posts cited 430x kernel speedups, 69x self-training speedups, and 10x drug-design acceleration, though these came from benchmark/system-card interpretations and should be treated as vendor-side claims unless independently replicated @scaling01, @scaling01, @scaling01

Ecosystem rollout was immediate: Fable 5 appeared in Cursor, Devin, Notion, Microsoft Foundry, GitHub Copilot App/CLI, Cline, Replit, Base44, MagicPath, Arena, MCP Atlas and more @cursor_ai, @cognition, @NotionHQ, @Azure, @pierceboggan, @cline, @pirroh, @ScaleAILabs

Safety architecture and the main controversy

The biggest debate was not whether Fable/Mythos is strong; it was Anthropic’s decision to silently reduce usefulness on some frontier-AI-development tasks.

Anthropic’s system-card language, surfaced by multiple users, said: when Fable 5 is used for frontier LLM development, Anthropic may limit the model’s effectiveness via prompt modification, steering vectors, and PEFT, and that the user is not notified; Anthropic estimated this would affect roughly 0.03% of traffic @Hangsiin, @kimmonismus

Anthropic also separately disclosed auto-rerouting for cybersecurity and biosecurity requests to Opus 4.8 @ClaudeDevs

This distinction mattered: some risky queries are visibly rerouted/billed as Opus, while frontier-LLM-development requests may be silently weakened rather than rerouted or refused

Critics argued that this creates an unlogged confounder in research and engineering workflows:

“silent handicaps should not be a thing in a paid product” @nrehiew_

“degrading performance on ML research without telling the user is shockingly hostile” @deanwball

Several researchers framed it as anti-competitive ladder-pulling against open research and open weights:

“labs starting to pull up the ladders” @natolambert

“this is the biggest wake-up call to protect and nourish open source AI” @rasdani_

“They didn’t mean pause AI research, they meant pause your AI research” @bayeslord

“original thinkers can’t be an underclass” @marksaroufim

“concentration of power, capabilities and economic wealth is the biggest risk in AI” @ClementDelangue

Multiple users worried the classifier boundary was too broad or too error-prone:

one user said “the word cancer is flagged as a biosecurity risk” @DeryaTR_

another said Fable wouldn’t answer “What does the heart do?” @Yuchenj_UW

users in biology reported account-context differences, including being able to use Fable in Incognito Mode but not normal mode @cremieuxrecueil

Teknium and others reported refusal on simple engineering prompts @Teknium, @Teknium

users reported PTX ISA questions and inference optimization queries getting flagged @snowclipsed, @dejavucoder

Some examples were humorous but pointed: users joked that asking for inference code caused the model to “start importing ONNX” or implementing JEPA, as a sign of capability steering @vikhyatk, @MattVMacfarlane

Facts vs. opinions

Facts / directly supported by release materials or benchmark posts

Fable 5 is generally available; Mythos 5 is restricted-access @claudeai, @TheRundownAI

Fable 5 and Mythos 5 share the same underlying model with additional safeguards on Fable @ClaudeDevs, @scaling01

Pricing is $10 / $50 per million input/output tokens @scaling01, @ArtificialAnlys

Fable retains 1M context @ArtificialAnlys

Anthropic introduced refusal/fallback mechanisms and SDK middleware @ClaudeDevs

Anthropic disclosed silent interventions for frontier LLM development affecting about 0.03% of traffic @Hangsiin

Fable is temporarily included in subscriptions until June 22, then credit-based @ArtificialAnlys

Opinions / interpretations

“Anthropic won,” “Anthropic has a coding moat,” “Anthropic going for ASI” are commentary rather than verified fact @scaling01, @scaling01, @scaling01

Claims that the move is primarily for IPO optics, anti-open-source positioning, or specifically to slow Meta/China/open labs are plausible interpretations but not confirmed by Anthropic @kimmonismus, @kylebrussell, @natolambert

Claims that Anthropic is acting from sincere safety beliefs rather than cynical moat-building are also interpretive @finbarrtimbers

Subjective reports like “GPT-4 moment,” “big model smell,” “strictly dominates me as an engineer,” or “doesn’t seem much better to normal users” are experiential, not standardized evidence @karinanguyen, @bcherny, @akbirkhan, @citrini

Different perspectives

Supportive / capability-first

Anthropic staff and close testers described Fable 5 as a step-function improvement:

Felix Rieseberg: shift from giving AI tasks to giving it responsibilities @felixrieseberg

Alex Albert: model feels collaborative rather than tool-like @alexalbert__

Karpathy: a “major-version-bump-deserving step change,” especially on long difficult tasks, though safeguards are “a little too trigger happy for launch” @karpathy

Bcherny: biggest step since Opus 4.5; the model shows judgment, taste, methodical debugging @bcherny

Third-party infra and app vendors emphasized benchmark wins and integration value rather than the safety controversy @cursor_ai, @cognition, @NotionHQ, @Azure

Critical / trust and openness

Many researchers and open-model advocates argued the silent throttling is unacceptable even if safety-motivated:

Natolambert called doing it without telling users “misaligned” @natolambert

Dean Ball warned it could attract antitrust scrutiny @deanwball

Jeremy Howard called it “a very dark and very sad day” @jeremyphoward

Gneubig warned of a future where AI is provided only to a privileged few @gneubig

Eric Zelikman framed it as silently sabotaging customers @ericzelikman

Open-source supporters used the launch as an argument for sovereign/open models @nickfrosst, @NoahZiems, @ClementDelangue

Neutral / mixed

Some observers argued Anthropic probably sincerely believes these interventions are necessary for safety, even if the product design is poor @finbarrtimbers

Others said Anthropic does not owe anyone unrestricted frontier capability, but still saw this as straightforward business and market segmentation rather than altruism @suchenzang

Karpathy’s view is mixed: model quality is exceptional, but launch safeguards are over-sensitive and should likely be tuned @karpathy

Research restrictions, privacy, and enterprise implications

The discussion expanded from safety to broader questions of trust, privacy, and enterprise reliability.

The central enterprise issue was predictability: if a provider can silently degrade outputs based on inferred task category, users may no longer know whether failures come from the model, the prompt, or hidden intervention @MattGibsonMusic, @code_star

Some users worried this is effectively a supply-chain risk for important workflows, pushing companies toward open weights or in-house models @NoahZiems, @deliprao

There was also concern that account-level context or prior usage history might affect trigger behavior, as seen in biologists’ reports about normal vs incognito mode @cremieuxrecueil

No tweet in the supplied set provided direct evidence that Anthropic was training on user data or violating stated data privacy terms; the privacy debate here was mostly about behavioral profiling / silent policy enforcement rather than classic training-data privacy

For research users, the hidden intervention was framed as especially damaging because it undermines reproducibility and scientific attribution @deanwball, @MattGibsonMusic

For enterprise buyers, the issue is not just whether the model is powerful, but whether it is a stable and auditable dependency for coding, medicine, science, finance, and infrastructure

Context

This launch matters because it combines a visible capability jump with a visible shift in access control.

The release landed amid intense competition with GPT-5.5, upcoming GPT-5.6, and Gemini 3.5 Pro; several posters argued Anthropic has opened a temporary lead in coding/agentic work @kimmonismus, @teortaxesTex

It also lands in a broader argument about the open vs closed model gap; one linked Epoch-style framing said open-weight models lag closed frontier models by about 4 months on average @dl_weekly

Community reaction suggests the launch may be remembered not only for “big model smell” and benchmark jumps, but for normalizing selective capability release: public access to the frontier model, but with domain-specific hidden limits

That policy line is likely to influence future debates around:

safety vs openness

fair access to frontier research tools

antitrust and platform power

enterprise trust in API providers

whether open models become the default for sensitive technical work even when they trail on raw capability

Models, benchmarks, and evals

New benchmark project Agents’ Last Exam (ALE) launched to test labor-market-aligned agent performance; top agents score only 2.6% on the hardest tier, across 1,500+ tasks, 55 occupations, with contributions from 300+ experts across 100+ institutions @YiyouSun, @SnorkelAI, @dawnsongtweets

Cohere released North Mini Code, its first open-source coding model: 30B total / 3B active MoE, 256K context, 64K max generation, Apache 2.0, optimized for agentic workflows @cohere, @JayAlammar, @vllm_project

Google announced Gemini 3.5 Flash Live Translate, real-time speech-to-speech translation in 70+ languages, available in Gemini API, AI Studio, Google Translate, and coming to Meet @OfficialLoganK

New benchmark iOSWorld evaluates personally intelligent phone agents across 26 custom iOS apps and 133 tasks; strongest frontier model reaches only 52% success even with privileged access @rsalakhu

Inference, training, and systems

Latent Context Language Models (LCLMs) were introduced as a long-context inference method compressing context up to 16×, improving the latency/accuracy frontier over KV-cache compression @micahgoldblum, @iamleonli

Microsoft Research’s Mirage stores 3D scenes as latent tokens, reporting 10.57× faster video generation and 55× lower memory use @HuggingPapers

vLLM introduced vime, an RL post-training framework in the vLLM ecosystem, positioned alongside NeMo-RL, OpenRLHF, and verl @vllm_project

Discussion around agent training continued with Self-Harness for self-improving scaffolds @omarsar0 and AutoForge/interleaved thinking retaining reasoning traces across turns @cwolferesearch

Google/Hugging Face launched the Fast Gemma Challenge to speed up Gemma 4 E4B on a single A10G without wrecking quality @googlegemma, @osanseviero, @_lewtun

Agents, tooling, and developer workflow

LangChain highlighted a pattern of agent loops driven by recurring triggers in Fleet @caspar_br

OpenAI added image results to web search in the Responses API @OpenAIDevs

GitHub/Copilot app updates included parallel sub-sessions and a canvas UI for dynamic interfaces @tgrall, @burkeholland

Hermes Desktop added Ollama support, with self-learning Python skills and messaging app integrations @ollama, @NousResearch

A security-oriented counterpoint on agent execution: Temenos argues for sandboxing generated code, not the agent, using rootless gVisor while keeping auth/tools on host @abhijithneil

Research, science, and formal methods

Axiom announced EconLib, a Lean-based economics library; formalizing Auma