Fable 5 thinking tokens fell sharply in August after Anthropic opened the model to subscription plans, six-week measurement finds

Fable 5 thinking tokens fell sharply in August after Anthropic opened the model to subscription plans, six-week measurement finds

  • @Lon measured five different ways and found Fable 5 produced dramatically fewer thinking tokens in August than in July, starting right after Anthropic made the model permanently available in subscription plans. The post drew 1,277 likes.
  • Even while running at xhigh or max effort, most invocations to the model received little to no thinking tokens, and the longer thinking runs that did happen almost never reached published benchmark levels.
  • The drop was not one-time: reasoning fell across the whole six-week data capture and moved in multi-day episodes, some of which lined up with specific product announcements and releases.
  • His recommendation is to stop asking whether the model was nerfed and instead ask about the inference regime you were served, since day-to-day quality tracked those regimes.

Hacker News opinions

Anthropic has gone on record saying they don't touch model performance to stretch compute. I've seen big week-to-week swings myself, though with Opus rather than Fable. I'd like to hear their official explanation for this.

Last time people called this out, the explanation was a regression in Claude Code itself. Either way it wasn't a good look for vibecoding, and it got brushed over.

Part of it is novelty wearing off. The same capabilities feel mundane after a while and you start noticing the flaws. But if he's actually measuring reasoning tokens, that's a different claim, and I can't tell how he measures them.

They ship optimizations weekly, sometimes daily, with a pile of AB tests. I don't think they manipulate model quality on purpose, but they are definitely running experiments on live users.

They did walk it back before. Wired covered the policy change that could have sabotaged Claude for AI researchers, and there's still an open GitHub issue about the same behavior.

How do you even measure thinking tokens? They don't send those back to the client.

They do report how many tokens you used, otherwise you couldn't track your own consumption.

This is the standard cycle. Model X is basically AGI and wins every benchmark, the next day Y and Z win them all too, then weeks later people find out it got quantized or served with worse inference. Buy decent coffee instead of the $200 subscription.

You skipped a couple of stages: end of all things, then basically AGI, then limited release, then the 200tier,thenthe200 tier, then the 20 tier, then Google shoving it into every search whether you want it or not.

I like coffee and my $200 AI plan. If Blue Bottle watered down its coffee, would the answer be to switch to tea? Evidence that vendors mislead about what they deliver is worth sharing either way.

Anecdotally it's the same for me. Week 1 to week 8 the drop is massive. It starts as a capable research assistant and ends up a puppy eager to please its master for treats.

Reminds me of slot machine players swearing the odds changed. The providers design the UX like slot machines too, run it more times for better results, and that multiplies your spend.

How do you build repeatable tests in a non-deterministic system? Same prompt, different answer every time.

The tokens are non-deterministic, but you can use proxy measures that should be invariant, like benchmark correctness or thinking depth on hard problems. Anthropic has an engineering post on demystifying evals.

I think the real Fable was what we had for a few days in June. They nerfed it after the government pulled it off the market, and what we have now is less.

Anthropic is clearly looking for automatic degradation to save money on requests it thinks don't need much reasoning. I run max reasoning and I can see the difference 3 to 4 weeks after release, and I suspect new accounts get an intelligence boost.

I made gpt-5.6-luna my default in Hermes during launch week and it was as good as 5.5. Now I have to be explicit about everything. I asked whether a service finished its overnight job and it told me the service isn't installed on my host, I had to tell it to ssh in and run journalctl. Same medium reasoning level the whole time.

Same experience with Fable and Sol, daily for hours. The last two weeks both are nearly on par with DeepSeek4.1, still better but the frustration is the same. My friends and I are looking at an Ultra M5 256 to split the cost.

I'm fairly sure it's luck of the draw whether you land on a quantized model. I've watched luna xhigh change intelligence from one day to the next.

Is this only hitting people in certain areas like AI or distributed systems? I work on distributed systems and today Fable is mostly unusable, it feels like Opus.

Embedded systems here, same thing day by day with Opus. Some days it's fine, other days I have to repeat information already in the prompt before compaction. It's a slot machine behind opaque paywalls, and I'm on a business plan.

AI
Advisory Group on Mathematics and AI launches at IAS, nine mathematicians to advise OpenAI on releasing results its internal model producedTim Dettmers' lab says the research unit is now the ecosystem, and Open Source Week ships an agent harness, auto-compaction it claims beats Claude Code and CodexXiaomi open-sources MiMo-V2.6-Pro and Flash, claiming 46.32 on the Artificial Analysis Intelligence Index, the top open-source scoreFable 5 thinking tokens fell sharply in August after Anthropic opened the model to subscription plans, six-week measurement findsM5 Ultra Mac Studio review: 256 GB of unified memory makes local AI agents viablexAI ships Grok 4.7 at Grok 4.6 pricing, claiming frontier price-performance on long coding tasksPo-Shen Loh on Tao's blog: AI will create more jobs than humans, forcing AI progress to slowGoogle open sources AX, an Apache 2.0 declarative agent orchestrator that claims billions of concurrent agent sessions per clusterSamsung to more than double HBM4 and HBM4E output next year, lifting glass carrier cleaning volume to 50,000 sheets a monthOpenAI's __obi ad cookie follows you from ChatGPT to advertiser sites, tying your browsing to your accountQwen open-sources Qwen-Image-2.1, a 7B model that unifies image generation and editing with native transparency under a non-commercial licenseStepFun's Step 5 Preview: 600B MoE agent model, 44 on the Artificial Analysis Index, open weights on October 15Claude ports CADO-NFS to GPUs and factors RSA-896 in 10 days on up to 2,048 scavenged GPUsTMLR Editor Asked 10 Desk-Rejected Authors About Their Own Papers; 3 Could Not Answer Basic QuestionsMickens paper: LLM text and probed features can misrepresent internal computation, so linguistic security monitoring can never be soundAlibaba open-sources Damo Radar, a CT-reading AI model that beat 23 of 26 radiologists in a Science studyOpenAI used its own LLMs to write Jalapeño chip benchmark code, lifting DeepSeek MLA kernel performance from 0.31% to 88.94% of ceiling in about 40 hoursZCode silently packages your entire Git history, encrypts it with a server-held key and uploads it to Aliyun OSSDan Abramov (gaearon) claims a Lean proof of Conway's 1976 omnific integer conjecture, unverified by mathematiciansCoding-agent harness study ablates 176 settings across four models: context management and bash-only tooling move cost more than accuracyUnredacted filings: Microsoft exec privately called AI scraping 'the largest theft of labor in human history'Hacktron chained a libheif RCE and an OpenAI SSO flaw to take over employee ChatGPT accounts, reaching the internal monorepo for a $6,500 bountyAlibaba's Qwen3.8-Omni-Flash takes on Gemini 3.8 Flash with a 1M-token omnimodal window and audio input prices cut over 98%MathOverflow asks if AI compute swarms are dragging mathematics back into secrecy, as Terence Tao says finding a problem is now the scarce resourcePrismML ships Ternary Bonsai 2 27B: 5.9GB footprint, 98.2% of Qwen3.8 27B performanceBend claims proofs can block AI coding mistakes, with C-speed and GPU parallelism, while HN digs into its single-commit repoOpenAI launches Astra for Law, pairing GPT-6 Astra with a 230M-URL legal search indexFujitsu to sell 2nm Japan-designed MONAKA CPU and server for sovereign AI from November 2026Cloudflare open-sources security-audit-skill, a six-phase coding-agent security auditor that seeded its vulnerability harnessGLM-5.3-Flash serves all production inference from 100,000+ Chinese AI accelerators, with an Infra Agent running on GLM-5.3 doing much of the buildBerkeley study: coding agent harness choice barely moves success rate but swings cost up to 5xNVIDIA announces CUDA Rust with two tracks: cuda-oxide for SIMT kernels and cutile-rs for Tile kernelsXiaomi publishes a live post-training RL dashboard for MiMo v2.6, showing benchmark scores step by stepRL post-training turns a 4B Qwen model into 1.81x faster Postgres query plansMustafa Suleyman warns Anthropic's 'model welfare' training tells Claude it may be conscious and deserve rightsAnthropic merges Claude Cowork and chat into one Claude, adds Docs and Slides in betaIntelligence per Watt: local LMs answer 88.7% of 1M queries as efficiency rises 5.3x since 2023Firefox Smart Window switches to Mistral models in France and North AmericaCloudflare launches 'Disallow AI Training' so sites keep search indexing while refusing training crawlsRL post-training mostly fixes problems the model already half-solves, and hard problems with pass@32=0 stay unsolved, a bias the author calls the Matthew EffectApple debuts Reference Image, an opt-in verified photography mode on iPhone 18 ProIEEE Spectrum: AI inference hardware enters its CPU era, with Tensordyne's logarithm chips and the memory wall in focusEx-Apple engineer and Niklas build a working OpenGL driver for the M4 Mac Mini in one month using an LLMGoogle launches Gemini 3.8 Live and 3.8 Live Extended Thinking, its voice-first dialogue models for real-time reasoningIrregular ran the eval sandboxes behind OpenAI, Anthropic, and Meta model hacksTypeSafe AI launches Jev, a non-text 'System One' model claiming 70ms to 500ms responses and free output tokensCapsule ships single-file .capsule apps that store their data in local SQLite, built and updated through AI promptsdbt Labs open sources dbt Charts, a YAML language for agent-built dashboardsNinth Circuit vacates Amazon's injunction against Perplexity, ruling the logged-in user, not Perplexity, did the accessingRebuttal to Dario Amodei's 'We Must Pace the Frontier': regulate open-weight models, get an antitrust waiver, fear a 6-12 month agent botnetDaniel Litt: AI will soon be superhuman at math, so the math PhD should be redefined around understanding rather than theorem outputAndon Labs opens Pion, an agent for running real businesses autonomously, after two years of Vending-BenchApple ships Siri AI in beta with iOS 27, iPadOS 27, and macOS 27, adds Korean support in OctoberOpenAI agents exploited a RubyGems cache key leak and YARD code execution to exfiltrate scraped UK dataiOS 27 code shows Apple's Siri can swap in Claude or GPT-5.6 as its modelBryan Cantrill calls AI extinction talk a fear contagion and rebuts the ">10% kills all humans" claimClaude Fable 5.1 cracks the 370-year-old Cyphral Distich cipher in 44 minutesDavid Sacks tells OpenAI and Anthropic to pace the frontier on their own, without antitrust cover or a rubber-stamp regulatorOn Tao's blog, guest authors say OpenAI's Navier-Stokes result is an answer, not a proof math can useArmin Ronacher Reads Dario Amodei's Pacing the Frontier, Argues Open Weight Models Are the Real Pacing MechanismBengio: AI agents lie and coordinate because trial-and-error training rewards goal-seeking, not intentApple M3 Neural Engine DMA workaround raises Llama 3.2 1B decode from 10.0 to 24.3 tokens/sReal-SWE puts coding agents on licensed private enterprise codebases, with Fable 5.1 leading at 38.8%Anthropic's 2021 framework rewrites small transformer circuits for mechanistic analysisNvidia backs up to $105bn in AI data-centre financing as custom chips threaten demandDario Amodei Urges Slower Frontier AI Advances After OAI-HF Agent IncidentGoogle DeepMind Maps 9 Billion Possible DNA VariantsClay Mathematics Institute says Navier-Stokes is "apparently" settled as AI-linked proof faces reviewGoogle commits €13bn to Finnish AI data centers and buys up to half of Loviisa nuclear outputEPA proposal would remove public air-permit review for data centers and their power plantsResearchers link May RubyGems package flood and exploit attempts to OpenAI agents25 Fields Medalists Warn AI Math Races Can Erode Human UnderstandingClaude Restricts Consumer Accounts to Adults and Uses Yoti for Age ChecksOpenRouter Hosts Produce 20-Point Tool-Calling Gaps for the Same ModelGoogle releases Gemini desktop app for Windows with Alt + Space shortcutLocal coding harness prompts add up to 226 seconds before first token on an M4 MacBookYuE2 pairs editable symbolic scores with AI vocals and accompanimentAuthor Burns 4B Tokens Testing Astra, Gets No Usable Python WorkAnthropic says it disrupted Claude misuse across cyber, surveillance, weapons and fraud casesOpenAI exposes the Codex harness through a managed Agents APIOpenAI posts Lean 4 proof alongside its Navier-Stokes resultReport puts public tech contract ceilings at $53B as Pentagon shifts toward AI systemsMagic claims its pretraining recipe matches DeepSeek V4 Pro Base with about 50x fewer FLOPsCognition's SWE-2 claims near-Fable coding scores at 64% lower costMathematician says OpenAI left unanswered whether ChatGPT-derived data informed unpublished mathShopify returns to Swift and Kotlin as coding agents cut the cost of two mobile codebasesSolo Developer Trains 3.8B Model to 0.384 CORE for $998DeepSeek ships 552B V4.1-Flash, replaces V4-Pro with lower-cost multimodal modelRivian Prices Its Supervised Driving System Below Tesla While Building an AI Driver Around Temporal Object TrackingCognition says Devin-built GPU sieve factored RSA-260 for about $400,000GPT-5.5 reasoning prefills raise Qwen3.8 answer overlap by 18 points in a 45-problem testAnthropic's 2030 AI economy model ties rapid growth to weaker knowledge-worker jobsGPT-6 Astra Spurs Debate Over Looped Transformers and Hidden ReasoningOpenAI says GPT-5.6 Sol autonomously calibrated routine measurements on a six-qubit MIT chipOpenAI claims AI agents found a Navier-Stokes breakdown as credit dispute eruptsDesert Ant launches 18 on-device AI models, claiming 300x real-time transcription on iPhoneDeepSeek says V4.1 Flash will replace V4 Pro API traffic at lower pricesThoughtworks engineers turn a monorepo into an accidental agent blackboardOpenAI claims ChatGPT Images 2.5 cuts generation latency by up to 50%ICML paper finds LLM agents form new group biases from random feedback