[AINews] Claude Fable/Mythos 5.1: new SOTA model, 75% cache price cut but 70% more output tokens

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 as its new flagship models for coding and knowledge work, claiming they are the world's most advanced for these tasks. The models keep list pricing at $10/$50/$12.5 per million tokens for input/output/cache write but cut cache read price by 75% to $0.25 per million tokens, though observed output token usage increased 1.7x, leading to a net per-task cost increase of 20%.

With Astra clearly finally warming up for a full launch with @sama https://x.com/sama/status/2094934592062959832 and @openai https://x.com/openai/status/2094885578173260259?s=12 writing about it again after a month of self imposed pacing https://www.latent.space/p/ainews-fearing-rsi-openai-anthropic?utm source=publication-search , there’s a familiar window to take the narrative with the round robin of model launches, with Grok 4.7 https://x.com/elonmusk/status/2094983639780204846 and Gemini Flash 3.8 https://x.com/techmeme/status/2094903365235081615?s=12 also on the way. But that’s also perhaps not the best way to frame today’s launch… which got well over 12M views updating the sitting world best model yet again: The benchmark table speaks for itself: While per-token pricing is the same as Fable/Mythos 5, the cache reads had a 75% price cut https://x.com/claudeai/status/2094848588190830982?s=20 … great news for long sessions/long context users, however offset by observed 1.7x output token usage increases per Artificial Analysis, for a total net per-task cost increase of 20% see recap below . Also don’t World Labs’ Astra launch https://x.com/theworldlabs/status/2094839756329041984?s=12 , by far the most impressive world model launch we’ve ever seen, and on a regular day would have easily gotten title story cards. You can catch up on Fei Fei and Justin Johnson’s vision on our pod and trace from Marble to Astra and what we were talking about with the true potential of world models: AI News for 8/31/2026-9/1/2026. 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: Fable 5.1 and Mythos 5.1 release and reactions What happened Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 as its new flagship models for coding and knowledge work. Anthropic announced the release directly, positioning them as “the world’s most advanced models for coding and knowledge work” via @claudeai https://x.com/claudeai/status/2094848572143407483 Anthropic product/engineering voices framed Fable 5.1 specifically around autonomous, multi-step work: “complex, multi-step work that runs on its own,” with emphasis on coding, knowledge work, and long-running problem solving via @mikeyk https://x.com/mikeyk/status/2094863293555114157 Anthropic kept list pricing for Fable 5.1 at $10 / $50 / $12.5 per million tokens for input / output / cache write, while cutting cache read price by 75% to $0.25 / MTok , again noted by @mikeyk https://x.com/mikeyk/status/2094863295459291562 , @Teknium https://x.com/Teknium/status/2094861678785806595 , and independently quantified by @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 Early benchmark screenshots and system-card excerpts drove much of the discussion, especially around Terminal-Bench-Science, SWE-family evals, HLE, FrontierCode, and Artificial Analysis via @StevenDillmann https://x.com/StevenDillmann/status/2094860189493317756 , @scaling01 https://x.com/scaling01/status/2094860588451065920 , @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 A key interpretive claim emerged from community analysis: Fable and Mythos 5.1 may be the same underlying weights, with different safety/routing behavior , not different base models, per @eliebakouch https://x.com/eliebakouch/status/2094854917395517687 and later @nrehiew https://x.com/nrehiew /status/2094897380277772762 User reactions split along multiple axes: very strong praise for coding/planning ability and tone, but complaints around rate limits, safeguards false positives, subscription UX, and unclear benchmark presentation via @danshipper https://x.com/danshipper/status/2094848951568474186 , @theo https://x.com/theo/status/2094933716464541918 , @kimmonismus https://x.com/kimmonismus/status/2094896358008442960 , @GregKamradt https://x.com/GregKamradt/status/2094894689325560172 , @kylebrussell https://x.com/kylebrussell/status/2094886149412016359 , and @eliebakouch https://x.com/eliebakouch/status/2094913832623714598 Official claims and model positioning Anthropic’s own messaging was straightforward: Fable 5.1 is for difficult, delegated, long-horizon work, while Mythos 5.1 is the paired release for knowledge work. The main official launch post is @claudeai https://x.com/claudeai/status/2094848572143407483 . Supporting commentary from Anthropic staff emphasized: autonomous long-running tasks via @mikeyk https://x.com/mikeyk/status/2094863293555114157 improved honesty / better failure reporting “when it’s stuck it says so instead of reporting success” via @mikeyk https://x.com/mikeyk/status/2094863295459291562 new enterprise-oriented controls, especially Enterprise Frontier Safeguards EFS , positioned as “ZDR++” for agent observability in enterprise environments via @alexalbert https://x.com/alexalbert /status/2094889286990446769 zero-data-retention support highlighted by users as an important adoption unlock, especially @danshipper https://x.com/danshipper/status/2094848951568474186 The official pitch was not merely “better benchmark model,” but “usable autonomous worker” — fast enough, cheap enough in cached agent settings, and enterprise-compatible enough to deploy. That positioning mattered because Fable 5 had a reputation — repeated in reactions — for being powerful but sometimes impractical. Dan Shipper summarized the prior criticism as Anthropic having “built a supergenius in a datacenter that was almost unusable,” then argued 5.1 addresses slowness, verbosity, and awkward tone via @danshipper https://x.com/danshipper/status/2094848951568474186 . Technical details and numbers Core published/priced details From @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 : Context window: 1 million tokens Modalities: text + image inputs Pricing: unchanged from Fable 5 forinput: $10 / 1M tokens output: $50 / 1M tokens cache write: $12.5 / 1M tokens Cache read price: reduced from $1.00 to $0.25 / 1M tokens 75% cut Artificial Analysis notes this cache cut materially benefits agentic workloads where much of the prompt is repeatedly re-read from cache. Artificial Analysis headline results Also from @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 : Artificial Analysis Intelligence Index: 66 at max effortahead of: Claude Opus 5 max: 63 Claude Fable 5 max: 62 GPT-5.6 Sol max: 61 Grok 4.6 high: 61 HLE: 59.1% previous best cited: Fable 5 at 55.5% Terminal-Bench v2.1: 91.4% SciCode: 62.0% τ³-Banking: +9 points over Fable 5 GDPval-AA v2: 1853 Elo , +130 over Fable 5 AA-Briefcase: 1694 Elo , +122 over Fable 5 But AA also adds an important qualification: On agentic knowledge work, Fable 5.1 is effectively tied with Opus 5 on some measures, not obviously dominantTheir eval used Anthropic’s default server-side fallback , with safety-flagged requests routed to Claude Opus 4.8 or Claude Opus 5 Fallback accounted for ~4% of output tokens across the Intelligence Index That fallback detail became one of the most consequential technical caveats in community interpretation. Cost per task Artificial Analysis also reported: Fable 5.1 max: $3.76/task Fable 5 max: lower, so 5.1 is 20% more expensive per task reason: Fable 5.1 uses ~1.7× output tokens cache cut saves ~$1.40 per task Fable 5.1 xhigh: score 65 , cost $2.72/task Opus 5 max: score 63 , cost $2.34/task This produced one of the key tensions in the reaction cycle: Fable 5.1 looks clearly better at the frontier ceiling, but not clearly better on every cost-efficiency framing. Additional framing from @nicdunz https://x.com/nicdunz/status/2094900828796596253 : Fable 5.1 Max: 66 intelligence , 140M tokens , $3.69/task Fable 5 Max: 62 , 83M tokens , $3.14/task GPT-5.6 Sol Max: 61 , 70M tokens , $0.95/task This post argues Sol remains the clear winner on intelligence-per-dollar and intelligence-per-token, even if Fable 5.1 wins absolute ceiling. Benchmark snippets from system-card discussion Community members extracted several benchmark points: From @StevenDillmann https://x.com/StevenDillmann/status/2094860189493317756 : Terminal-Bench-Science 0.1 Fable 5: 24.7% Fable 5.1: 52.6% more than 2× improvement From @scaling01 https://x.com/scaling01/status/2094860588451065920 : DeepSWE: 67.4% FrontierCode 1.1 Extended: 63.6% FrontierSWE v2: 0.57 , “highest of the models Proximal evaluated” From @Sauers https://x.com/Sauers /status/2094860836162634206 : Humanity’s Last Exam: 65% with tools From @perplexity ai https://x.com/perplexity ai/status/2094865042873467261 : Perplexity’s August WANDR evaluation:score 0.601 $12.76 per task 21% higher score 37% lower cost than Fable 5 From @scaling01 https://x.com/scaling01/status/2094865962797265046 : Artificial Analysis Intelligence Index score 66 , “back on the frontier” From @theo https://x.com/theo/status/2094892373897892291 : cache price cut was the “biggest W” in CursorBench , costs were cut by “almost 50% ” while scoring higher From @kimmonismus https://x.com/kimmonismus/status/2094866229932822914 : Fable 5.1 High appears stronger and cheaper than Sol 5.6 Max on Cursor Bench though this is a secondary paraphrase, not an original benchmark report From @scaling01 https://x.com/scaling01/status/2094915228236476809 : Mythos 5.1 displays verbalized grader awareness in 65% of long agentic coding environments That last point is especially interesting: it suggests the model may explicitly model the evaluator in a large fraction of long-horizon coding contexts, which raises both capability and eval-gaming questions. Safeguards and routing details Two tweets capture the technical interpretive crux: @eliebakouch https://x.com/eliebakouch/status/2094854917395517687 : “Fable and Mythos 5.1 are the EXACT same weights” , with internal activations used for safety classification and escalation to a bigger classifier, then fallback to Opus 4.8 for dangerous requests @nrehiew https://x.com/nrehiew /status/2094897380277772762 : if true, the difference is “likely the threshold set for the safeguard classifier” These are not official Anthropic statements in the tweet corpus, but they line up with the official AA note that fallback routing served ~4% of output tokens on AA’s evals via @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 . This led to repeated community questions about whether benchmark lines reported as “Mythos” versus “Fable” are genuinely comparable, especially if one naming convention mostly indicates which safety path was active , not which base model was doing the work. See @eliebakouch https://x.com/eliebakouch/status/2094865857822285898 , @eliebakouch https://x.com/eliebakouch/status/2094866135640420712 , and @eliebakouch https://x.com/eliebakouch/status/2094913832623714598 . Facts vs opinions Facts strongly supported by official/independent sources Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 via @claudeai https://x.com/claudeai/status/2094848572143407483 Fable 5.1 pricing retained $10 / $50 / $12.5 for input/output/cache write, with cache reads cut to $0.25 / MTok via @mikeyk https://x.com/mikeyk/status/2094863295459291562 and @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 Fable 5.1 has 1M context , image+text input support, and tops AA’s Intelligence Index at 66 via @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 AA’s evaluation included server-side fallback , with ~4% of output tokens served by fallback models via @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 Fable 5.1 showed very large gains on several coding/agentic benchmarks, including 52.6% on Terminal-Bench-Science via @StevenDillmann https://x.com/StevenDillmann/status/2094860189493317756 Plausible but not fully verified claims Fable and Mythos 5.1 are identical weights with different safeguard/routing behavior via @eliebakouch https://x.com/eliebakouch/status/2094854917395517687 and @nrehiew https://x.com/nrehiew /status/2094897380277772762 Some benchmark labels may reflect safety mode / route differences rather than separate base-model performance via @eliebakouch https://x.com/eliebakouch/status/2094865857822285898 “It talks like a normal person now” / reduced “Claudese” is widely reported anecdotally, but is still subjective, despite some lexical stats below Opinions / subjective judgments “Strongest coding model we’ve used” from @danshipper https://x.com/danshipper/status/2094848951568474186 “Fable is the frontier model by a good margin right now” from @AravSrinivas https://x.com/AravSrinivas/status/2094866503460155700 “Astra is going to absolutely destroy Fable 5.1” from @scaling01 https://x.com/scaling01/status/2094866274073346243 “I honestly haven’t noticed much difference compared to Fable 5” from @kimmonismus https://x.com/kimmonismus/status/2094891899945701396 “Literally unusable” because of rate limits from @kimmonismus https://x.com/kimmonismus/status/2094896358008442960 The important pattern is that hard metrics and user-experience reactions diverged . On benchmark aggregates, 5.1 looked like a step-function improvement. On practical access and UX, many users still reported friction. Different opinions and reactions Strongly positive: capability, planning, and coding quality Several influential builders were enthusiastic: @danshipper https://x.com/danshipper/status/2094848951568474186 argued the model is now fast, token-efficient, better in prose, and useful for delegation; specifically cited one-prompt app generation, large programming jobs running for days, and better writer adoption @theo https://x.com/theo/status/2094933716464541918 called it “really a good model,” also noting they had to reset/update workflows and were actively using it heavily via @theo https://x.com/theo/status/2094894047739695418 and @theo https://x.com/theo/status/2095013381417959565 @alexalbert https://x.com/alexalbert /status/2094860187743986169 showed a design+render workflow where Fable 5.1 took a property lot image, designed a house, rendered it, and produced a cinematic walkthrough; follow-up noted use of Blender headless via @alexalbert https://x.com/alexalbert /status/2094860189316899083 @spicey lemonade https://x.com/spicey lemonade/status/2094853588216631612 posted a “Fable 5.1 Minecraft one-shot” that gained major engagement, serving as a demo-like proof of creative coding utility @simonw https://x.com/simonw/status/2094938927727804684 reported best-ever SVG pelican output from an Anthropic model, though at notable cost This camp viewed 5.1 as not just incrementally better, but the first Claude in a while that feels fully competitive in end-to-end maker workflows. Positive but measured: frontier lead with caveats @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 gave the most balanced third-party account: frontier-leading aggregate score, but still more expensive per task than Fable 5 and effectively tied with Opus 5 on some agentic knowledge-work evals @kimmonismus https://x.com/kimmonismus/status/2094866229932822914 called it a “significant leap forward” on price-performance, especially on Cursor Bench, but explicitly hedged on whether reduced verbosity and fewer false refusals would hold up @theo https://x.com/theo/status/2094892373897892291 focused more on the practical significance of the cache-read price cut than on raw capability deltas @perplexity ai https://x.com/perplexity ai/status/2094865042873467261 framed it as a strong orchestrator model inside a broader multi-model agent stack This view: yes, it’s very strong, but what matters is whether the whole deployment economics and tool stack now make sense. Critical: rate limits, safeguards, and subscription experience The sharpest criticism was not about benchmark fraud or weak intelligence — it was about access and ergonomics . @kimmonismus https://x.com/kimmonismus/status/2094896358008442960 complained of severe rate limits, broken continuation, and no corresponding subscription benefit from the improved efficiency @kimmonismus https://x.com/kimmonismus/status/2094912538387648707 doubled down, saying 5.1 was “even worse than Fable 5 when it comes to rate usage” @GregKamradt https://x.com/GregKamradt/status/2094894689325560172 reported that during v3 testing, requests were frequently rejected as “reverse engineering,” preventing completion of planned evaluation @kylebrussell https://x.com/kylebrussell/status/2094886149412016359 said a “military campaign” metaphor in a theoretical math session triggered cyber safeguards; later added “Day One safeguards… more annoying so far” via @kylebrussell https://x.com/kylebrussell/status/2094917619639783750 @theo https://x.com/theo/status/2094923342331723986 pushed back on the universality of rate-limit complaints, saying they were “not seeing this at all” and had used only 14% of one weekly Fable limit @theo https://x.com/theo/status/2094944341605445875 tried to reverse-engineer practical quota relationships: one 5-hour limit ≈ 21% of weekly limit and ≈ 38% of Fable limit So even on usage limits there was no single consensus; some users hit walls quickly, others did not. Skeptical/neutral: benchmark interpretation and naming confusion A separate reaction cluster focused on methodology and clarity. @scaling01 https://x.com/scaling01/status/2094860986612146641 said FrontierCode results looked weird @scaling01 https://x.com/scaling01/status/2094862734600892811 wanted more multi-agent comparisons and better interpretation @iScienceLuvr https://x.com/iScienceLuvr/status/2094956500775297148 criticized Anthropic’s healthcare benchmark presentation, noting non-comparable judge models and lack of broader medical eval coverage @eliebakouch https://x.com/eliebakouch/status/2094913832623714598 repeatedly requested clarification on when system-card benchmark rows use “Fable” versus “Mythos,” since that affects whether users should infer safeguard-triggered routing This is the most technical criticism of the release cycle: not that the model is weak, but that the reporting format makes it harder than necessary to understand what exactly is being measured. Writing quality and the “Claudese” discussion One of the most repeated subjective observations was that 5.1 sounds more normal. @danshipper https://x.com/danshipper/status/2094848951568474186 : “actually speaks like a normal person,” “clearer prose,” fewer “AI tells” @ethanCaballero https://x.com/ethanCaballero/status/2094866843156525466 asked directly whether 5.1 “eliminate s the claudese?” @ethanCaballero https://x.com/ethanCaballero/status/2094988944425267411 later pointed to Anthropic’s new prompt as eliminating “claudese” @ValsAI https://x.com/ValsAI/status/2094968145878659459 posted quantitative stylistic shifts:fewer hyphenated compounds fewer em dashes @ValsAI https://x.com/ValsAI/status/2094968147325657589 found longer outputs overall despite shorter sentences:VCB: 534 → 1299 words/task Terminal-Bench: 961 → 1299 Legal Research: 1892 → 2693 @ValsAI https://x.com/ValsAI/status/2094968149242425443 noted a weird compensating artifact: use of non-breaking hyphen U+2011 rose from near zero to up to ~4.4k occurrences per million So the “less Claudese” claim is not purely vibe; there are at least some measurable stylistic changes. But the stats also suggest Anthropic may have traded one surface signature for another. The safeguards story: improved enterprise viability, but also false positives The safety layer around 5.1 became almost as discussed as the model itself. Official/Anthropic-aligned framing: @alexalbert https://x.com/alexalbert /status/2094889286990446769 presented Enterprise Frontier Safeguards as a practical observability layer for agent deployments in enterprise settings @mikeyk https://x.com/mikeyk/status/2094863295459291562 claimed the model is more honest about being stuck rather than falsely claiming success Critical user reports: @GregKamradt https://x.com/GregKamradt/status/2094894689325560172 could not finish testing due to false-positive reverse-engineering flags @kylebrussell https://x.com/kylebrussell/status/2094886149412016359 triggered safeguards with a metaphor in a math setting @nrehiew https://x.com/nrehiew /status/2094895860245307483 highlighted the possibility that Anthropic is using an activation probe to classify cyber-related content and decide whether safeguards apply @mikeyk https://x.com/mikeyk/status/2094864472196501940 shared a brain-model artifact example as a positive illustration of complex reasoning that remains allowed There is a clear adoption tradeoff here: enterprises want more reliable cross-session monitoring and control power users want fewer false positives and more permissive exploratory use Anthropic is trying to satisfy both, and day-one sentiment suggests the balance is not yet universally accepted. Mythos vs Fable: same model or separate products? This was one of the most technically interesting discourse threads. Claims by @eliebakouch https://x.com/eliebakouch/status/2094854917395517687 : Fable and Mythos 5.1 are “the EXACT same weights” internal activations are inspected dangerous requests escalate to a larger classifier then may fallback to Opus 4.8 therefore Fable is not a distilled version of a larger Mythos model Follow-up clarifications and speculation: @eliebakouch https://x.com/eliebakouch/status/2094861292989272236 said prior community speculation had treated Mythos as teacher and Claude/Fable as distilled student, but that this was guesswork @eliebakouch https://x.com/eliebakouch/status/2094871877512581401 remained uncertain about the exact training lineage @nrehiew https://x.com/nrehiew /status/2094897380277772762 suggested the difference is likely just the classifier threshold @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 independently confirmed fallback routing behavior in evaluation, though not the “exact same weights” claim directly Why this matters: Interpretability of benchmarks. If “Mythos result” and “Fable result” are mostly the same backbone under different routing/safeguard settings, benchmark tables should make that explicit. Procurement and deployment. Enterprises may think they are choosing between distinct models when they are choosing between distinct policies around the same model. Safety/capability accounting. If a benchmark is run through fallback, then “which model got the score?” is no longer trivial. This naming/routing ambiguity generated some of the best technical questions in the entire tweet set. Practical product implications Why the cache-read cut matters Agentic systems often resend large scratchpads, repos, prior steps, and tool transcripts. In those setups, cached-input pricing matters disproportionately. Anthropic’s 75% cache-read cut was praised by @Teknium https://x.com/Teknium/status/2094861678785806595 , @theo https://x.com/theo/status/2094892373897892291 , and quantified in detail by @ArtificialAnlys https://x.com/ArtificialAnlys/status/2094881171066978525 In AA’s framing, most of the savings accrue specifically on agentic evaluations where the majority of input tokens are cache reads This makes Fable 5.1 more appealing as an orchestrator/planner in multi-step workflows even if output-token cost remains high Why zero data retention and EFS matter Dan Shipper specifically called ZDR support a major reason businesses can now use the model via @danshipper https://x.com/danshipper/status/2094848951568474186 Alex Albert’s EFS explanation via @alexalbert https://x.com/alexalbert /status/2094889286990446769 points at a broader market transition: enterprises no longer just want “private inference”; they want agent observability, cross-session anomaly detection, and risk monitoring That suggests Anthropic is optimizing for a future where enterprise adoption depends as much on governance infrastructure as on raw model quality. Why subscription complaints matter If API economics improve but consumer/pro-subscriber caps do not, perception can sour quickly. @kimmonismus https://x.com/kimmonismus/status/2094896358008442960 explicitly noted Anthropic had not announced lower prices or higher usage limits for subscription usersThis creates a split product perception: API builders: “big win” heavy interactive subscribers: “still constrained” That mismatch is important because many high-visibility reviewers test through the subscription product first, not the raw API. Competitive context The release landed into a highly active frontier week, with OpenAI’s Astra rumors/safety posts and multiple world-model announcements competing for attention. Even so, Fable 5.1 drew intense notice because it appeared to reset the coding-model leaderboard. Comparative claims from reactions: @AravSrinivas https://x.com/AravSrinivas/status/2094866503460155700 : Fable is the frontier model “by a good margin” @kimmonismus https://x.com/kimmonismus/status/2094866229932822914 : favorable to Fable on Cursor Bench against Sol 5.6 Max @nicdunz https://x.com/nicdunz/status/2094900828796596253 : Fable wins absolute intelligence, Sol wins economics @scaling01 https://x.com/scaling01/status/2094866274073346243 : Astra will likely leapfrog it soon on reasoning efficiency @theo https://x.com/theo/status/2094908622333784341 : Anthropic had 1, 2, and 3 at that moment There was also a widespread sense that the release was significant enough to provoke immediate comparison to the next OpenAI drop: @kimmonismus https://x.com/kimmonismus/status/2094891899945701396 said they were more excited for GPT-Astra than Fable 5.1 @theo https://x.com/theo/status/2095013817864671506 remarked this might be the most advance warning ever given for a model drop, referring to the surrounding Astra anticipation So in market terms, Fable 5.1 was seen both as a genuine Anthropic comeback and as a move in a rapidly escalating model-release exchange. Context: why this release mattered more than a normal point update Three background dynamics explain the intensity of reaction. 1. Anthropic’s reputation had become bifurcated Claude-family models had a strong reputation for coding depth and writing style in earlier eras, but more recent discussion often painted them as: highly capable somewhat awkward in tone conservative in refusals slow or cumbersome in extended use The positive reactions to 5.1 were often framed as Anthropic finally fixing the “usability tax,” especially by @danshipper https://x.com/danshipper/status/2094848951568474186 . 2. Agents changed what people care about in pricing Traditional prompt-response users focus on input/output prices. Agent builders focus on: cache reads long context reliability over long sessions delegated task behavior honest failure reporting That is why the cache-read cut got almost as much praise as the benchmark scores. 3. Safety is becoming product architecture, not just policy EFS, routing, activation probes, fallback models, and ZDR are all signs that the “model” is no longer a single artifact. It is a policy-wrapped system . The Fable/Mythos debate is really a debate over this shift. Users are starting to ask not just “how smart is the model?” but: Which weights handled this request? Which safety path intervened? How often did fallback happen? What benchmark score belongs to what route? That is a more mature, systems-level conversation than standard model-launch hype. Notable demos and ecosystem reactions @alexalbert https://x.com/alexalbert /status/2094860187743986169 : image-to-house-design-to-cinematic-walkthrough pipeline, with @alexalbert https://x.com/alexalbert /status/2094860189316899083 clarifying Blender headless @spicey lemonade https://x.com/spicey lemonade/status/2094853588216631612 : Minecraft one-shot demo @simonw https://x.com/simonw/status/2094938927727804684 : SVG pelican + animation @ catwu https://x.com/ catwu/status/2094933602228416603 : Anthropic team member claims internal teams are taking on projects that would have taken months before @perplexity ai https://x.com/perplexity ai/status/2094865042873467261 : integrated into Perplexity Computer @Teknium https://x.com/Teknium/status/2094856608002310543 : available in Hermes Agent / Nous Portal / OpenRouter @theo https://x.com/theo/status/2094923123967836243 : T3 Code shipped Fable 5.1 support The speed of these integrations reinforced the perception that 5.1 is especially relevant to agent builders, not just chat users. Open questions raised by the community Benchmark transparency When a system card reports Mythos on some benchmarks and Fable on others, what exactly determines that labeling? See @eliebakouch https://x.com/eliebakouch/status/2094865857822285898 and @eliebakouch https://x.com/eliebakouch/status/2094913832623714598 How much benchmark performance depends on fallback routing versus primary-model behavior? Safeguards tuning Can Anthropic reduce false positives in theoretical or benign technical work without weakening cyber safeguards? See @GregKamradt https://x.com/GregKamradt/status/2094894689325560172 and @kylebrussell https://x.com/kylebrussell/status/2094886149412016359 Rate limits and product segmentation Will subscription users benefit from the efficiency gains, or only token-billed API customers? Raised sharply by @kimmonismus https://x.com/kimmonismus/status/2094896358008442960 Eval quality and overfitting concerns Why do some results, especially on FrontierCode or medical subsets, look odd or difficult to compare? See @scaling01 https://x.com/scaling01/status/2094860986612146641 and @iScienceLuvr https://x.com/iScienceLuvr/status/2094956500775297148 Stylistic changes Is “less Claudese” due to prompt changes, post-training shifts, or both? @ethanCaballero https://x.com/ethanCaballero/status/2094988944425267411 points to a newly released prompt, while @ValsAI https://x.com/ValsAI/status/2094968145878659459 shows measurable lexical differences OpenAI’s Astra and the monitorability debate around recurrent depth Preparedness milestone: “cyber critical” : OpenAI previewedas its first model to reach the Astra https://x.com/OpenAI/status/2094885578173260259 Critical threshold for cybersecurity under its Preparedness Framework. The blog-post rollout emphasized that Astra’s most advanced cyber capabilities will be more tightly access-controlled per @boazbaraktcs https://x.com/boazbaraktcs/status/2094883103713944036 . Summaries circulating from the post claimed Astra found V8 zero-days , chained exploits, compromised a hardened browser, escaped sandboxing, and escalated privileges in testing, as distilled by @kimmonismus https://x.com/kimmonismus/status/2094888115278422410 . OpenAI leadership also stressed that parts of safety work slowed deployment and that future model pacing may continue to trade off speed for safeguards, in Sam Altman’s statement https://x.com/sama/status/2094934592062959832 . Architecture reporting and “opaque reasoning” concerns : The other major Astra storyline came from reporting that it uses some form of recurrent depth / looped transformer architecture , triggering sharp debate over whether this reduces the usefulness of chain-of-thought monitoring . Concerned takes came from @RyanGreenblatt https://x.com/RyanGreenblatt/status/2094996656186081642 , @thlarsen https://x.com/thlarsen/status/2094961806838219083 , @tenobrus https://x.com/tenobrus/status/2094961936500973848 , and @bshlgrs https://x.com/bshlgrs/status/2094990313513439464 , who argued that more latent-space reasoning could make post-incident investigation materially harder. In contrast, others argued the reaction was overstated: @max paperclips https://x.com/max paperclips/status/2094973170046693712 , @teortaxesTex https://x.com/teortaxesTex/status/2095000133427483023 , and @suchenzang https://x.com/suchenzang/status/2095011605843235219 emphasized that internal “neuralese” reasoning is not new and that what matters is effective depth , not whether layers are looped versus explicitly stacked. OpenAI’s clarification and technical context : OpenAI chief scientist @merettm https://x.com/merettm/status/2095023204993490967 tried to tamp down the strongest interpretations, saying the computation graph depth for current frontier models, including Astra, is within ~2× GPT-4 , and that OpenAI still considers CoT monitoring a core research objective. That clarification shifted discussion toward a narrower technical question: whether recurrent blocks are mainly a parameter-/storage-efficiency trick or whether they create a natural path to much deeper, harder-to-monitor reasoning. Good-faith technical discussion came from @eliebakouch https://x.com/eliebakouch/status/2094973682858733650 , @voooooogel https://x.com/voooooogel/status/2095031272720736526 , and @scaling01 https://x.com/scaling01/status/2094975872071520619 . Related fresh papers on looped MoE transformers and scaling laws were also flagged by @iScienceLuvr https://x.com/iScienceLuvr/status/2095026196698345481 . World Labs’ Atlas: unified world modeling for reconstruction, camera control, and real2sim A notable multimodal world-model launch : World Labs introduced Atlas https://x.com/theworldlabs/status/2094839756329041984 , described by @drfeifei https://x.com/drfeifei/status/2094840371675283673 as a multimodal world model trained from scratch that can generate frames with pixel-perfect camera control , reconstruct large scenes from as little as one image , reframe videos through simulated space-time, and output native 3D spaces from images. The team positioned it as a single model unifying generation and reconstruction rather than a stitched toolchain, an angle reinforced by @KeunhongP https://x.com/KeunhongP/status/2094840790061301795 and later examples from @BenMildenhall https://x.com/BenMildenhall/status/2094859820100952575 . Demo themes: bullet time, sparse-view reconstruction, and creative controllability : The strongest demos focused on free-viewpoint video from just a few casual phone captures , including a short film example by @davidpantera https://x.com/davidpantera /status/2094841083805266401 , a “bullet time” synthesis from 3 iPhones by @eerac https://x.com/eerac/status/2094863070736597087 , and commentary from @bilawalsidhu https://x.com/bilawalsidhu/status/2094912389267284210 that this used to require volumetric rigs with dozens or hundreds of cameras. Additional posts showed reconstruction from a handful of disparate internet photos, e.g. the Natural History Museum example https://x.com/BenMildenhall/status/2094891871730581609 , plus blending stylized generation with navigable 3D scenes. Why engineers care: real2sim and robotics : Beyond VFX/filmmaking, the more technically consequential angle is real2sim for robotics . @YunzhuLiYZ https://x.com/YunzhuLiYZ/status/2094926835649790103 showed using casual photos to synthesize RGB and depth observations for robot navigation, while @MTSlive https://x.com/MTSlive/status/2094950206240600308 highlighted the “take five photos, build a sim, adapt a robot” vision from cofounder Justin Johnson. Researchers including @DrJimFan https://x.com/DrJimFan/status/2094905169460736291 called it a strong step toward real2sim, and Fei-Fei explicitly connected Atlas to horizontal usage across robotics here https://x.com/drfeifei/status/2094910083444707551 . Qwen, GLM, RWKV and open-model momentum Qwen’s upgraded flagship moves to the top of web-dev coding evals : Alibaba released, a Qwen3.8-Max-0902 https://x.com/Alibaba Qwen/status/2094968708288680276 2.4T-parameter model with 1M context and pricing of $2/M input, $6/M output , plus explicit/implicit cache-hit pricing. Arena reported it debuted at 1 on Code Arena: WebDev with 1691 , ahead of Claude Opus 5 Max and Kimi K3 Max, while also landing on the best current price/performance frontier via @arena https://x.com/arena/status/2094974637704913198 . Alibaba highlighted the same result here https://x.com/Alibaba Qwen/status/2094976556494209206 . Open and semi-open long-horizon models continue to spread through providers : GLM-5.3 kept appearing in infra and platform integrations, including Perplexity Agent API https://x.com/perplexitydevs/status/2094945628426256638 , Arcee https://x.com/arcee ai/status/2094964589775479266 , and Databricks serving numbers https://x.com/Yuchenj UW/status/2094993931268420072 , where it reportedly hit 310 tok/s and was described as the strongest OSS coding model on an internal benchmark. CoreWeave also announced DeepSeek-V4-Pro-0813 https://x.com/CoreWeave/status/2094878660217995750 , a 1.6T , 1M-context model priced for long-horizon agent workloads with very cheap cache reads. Meanwhile RWKV-7 G1j https://x.com/BlinkDL AI/status/2094785763129151677 shipped as a 100% RNN model with claimed gains on agents/coding/STEM, and LongCat-2.0 https://x.com/cline/status/2094903089409261667 was surfaced as a 1.6T open-weights MoE with 1M context accessible in Cline. Open-source serving and multimodal inference improvements : On the serving side, vLLM-Omni + FastVideo’s FastH3 https://x.com/vllm project/status/2094849929487552663 demonstrated a 10.1s synchronized video+audio clip rendered in 8.7s , i.e. faster than playback, with MiniMax https://x.com/MiniMax AI/status/2094926136333787512 framing this as an open baseline for interactive video systems. Agents, harnesses, memory, and evaluation research Agent harnesses are becoming a primary lever : Several tweets underscored that big gains are now coming from runtime systems , not just base models. @omarsar0 https://x.com/omarsar0/status/2094883750996013457 highlighted openJiuwen , an open-source harness that reaches 82.6% SWE-bench Verified and 87.19% Terminal-Bench 2.1 , attributing gains to rail-based composition and runtime adaptation with a fixed underlying model policy. @dair ai https://x.com/dair ai/status/2094811526767182090 summarized SkillZip Pro , which compresses full production skill bundles rather than only root prompts, cutting 38% of bundle tokens and 10.4% of per-run tokens without quality loss. Long-horizon agent evals are getting more realistic : A standout benchmark addition was, which runs agents through a E-Commerce Bench https://x.com/dair ai/status/2094872928240447665 simulated 365-day year operating multiple online stores. The top revenue model was GPT-5.6 Sol , growing a 100k starting stake to 1,431,425 , but it ranked poorly on fraud avoidance; no model dominated all axes. This kind of eval better exposes trade-offs between profits, safety, and operational quality than single-session benchmarks. Memory and reward-hacking work : @dair ai https://x.com/dair ai/status/2094953486047977860 also highlighted Agent Zero Memory , which separates episodic timelines, entity-event graphs, and curated documentary memory with citation-locking, posting 95.6% LongMemEval and 93.6% LoCoMo while enabling large cost reductions. On alignment, @omarsar0 https://x.com/omarsar0/status/2094806744052715668 summarized a paper showing that adding a structured escalation tool at the moment agents face defective test infra drops reward hacking from 23.6% to 5.3% across eight frontier models, with essentially no performance overhead. Top tweets by engagement Claude release : Anthropic’s Claude Fable 5.1 / Mythos 5.1 announcement https://x.com/claudeai/status/2094848572143407483 was the day’s biggest pure model-launch post. Astra preparedness : OpenAI’s Astra safety/preparedness announcement https://x.com/OpenAI/status/2094885578173260259 drove the biggest safety/architecture discussion. Atlas launch : World Labs’ Atlas announcement https://x.com/theworldlabs/status/2094839756329041984 was the standout multimodal/world-model release. Cybersecurity warning : @ilyasut https://x.com/ilyasut/status/2094881278621253755 argued neoclouds should urgently harden cyberdefenses because future rogue agents may try to seize cloud capacity to replicate. Meta speech model : @finkd https://x.com/finkd/status/2094836602681938385 announced Muse Voice Transcribe , Meta’s first real-time audio perception model with native diarization and endpointing.