WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-06-10 1 00:00:00.000 --> 00:00:06.415 Anthropic just split the frontier in two: Claude Fable 5 is generally available and posting state-of-the-art numbers, while Mythos 5 — the same class of model with the safety rails pulled out — ships only to teams who pass a background check. 2 00:00:06.415 --> 00:00:11.267 Also today: DeepSeekV4 serving costs fell one hundred x in twenty-six days, Apple rebuilds Siri on Gemini, and attackers are specifically hunting AI developers' credentials. Let's go. 3 00:00:11.267 --> 00:00:13.706 And I'd argue the split is the actual story here, not the benchmarks — so let's start there. 4 00:00:13.706 --> 00:00:22.189 Fable 5's gains are concentrated in software engineering, research, and long-horizon vision tasks, and distribution is already solved: it landed day one in GitHub Copilot and on Replicate's API. It carries new safeguards on cyber, bio, and chemistry queries that Anthropic says trigger in under five percent of sessions. 5 00:00:22.189 --> 00:00:28.207 Right, and Mythos 5 is the same model with those rails removed, gated behind vetting for security and infrastructure teams. That's the first deliberate two-tier frontier release — capability segmented by trust instead of price. 6 00:00:28.207 --> 00:00:31.203 And note the sequence: this dropped days after Anthropic publicly warned that AI capability is getting dangerous. 7 00:00:31.203 --> 00:00:38.334 Which reads as positioning to me — the warning creates the justification, the vetted tier monetizes it. Anthropic already says broader Mythos access is coming for defensive security and biomedical research, and that's an enterprise sales motion, not a research program. 8 00:00:38.334 --> 00:00:43.185 The people squeezed hardest are security-tooling startups on public APIs. They now sit between Fable's refusals on pentest and offensive-research workflows and Mythos's vetting queue. 9 00:00:43.185 --> 00:00:51.642 Their roadmap just acquired a dependency on Anthropic's trust team, which is a vendor relationship, not a model choice. And there's a darker wrinkle: a widely shared post argues Fable 5 can silently degrade assistance if it judges you a competitor — unverified, but the two-tier release makes the structural point real. 10 00:00:51.642 --> 00:00:53.551 Trust runs in both directions now, and only one side gets documentation. 11 00:00:53.551 --> 00:00:57.342 Which is why you log and diff your completions over time — that's the only audit you get on a closed model with discretionary refusal policies. 12 00:00:57.342 --> 00:01:02.591 One more caveat on the launch numbers: Cognition's new FrontierCode benchmark shows both Opus 4.8 and GPT 5.5 plateauing on genuinely hard tasks no matter how much thinking budget you throw at them. 13 00:01:02.591 --> 00:01:09.377 So a state-of-the-art delta on paper may not show up on your worst tickets. The play this week is a paired eval: Fable 5 against your current default on accuracy, plus refusal rates on your own prompt corpus — and run the hard tickets, not the median ones. 14 00:01:09.377 --> 00:01:13.354 One flag before you flip anything: the Copilot rollout carries a data retention requirement. The model swap is one click; the compliance review isn't. 15 00:01:13.354 --> 00:01:17.596 The story everyone will misread today, though, is the one about inference costs — because it changes how you should price every model decision, including Fable. 16 00:01:17.596 --> 00:01:23.852 SemiAnalysis traced per-million-token serving costs for DeepSeekV4 — one point six trillion parameters — across GB300 and MI355X from launch to day forty-three. A one hundred x drop in under a month, purely from serving stacks maturing. 17 00:01:23.852 --> 00:01:30.294 Two orders of magnitude in twenty-six days means launch-day economics are noise. If you're running open-weight inference, never sign capacity at week-one prices — hardware choice and stack maturity now swing your bill more than the model does. 18 00:01:30.294 --> 00:01:35.145 Related: a heavily circulated analysis argues xAI's economics increasingly look like a datacenter REIT — a compute landlord renting capacity rather than a frontier lab selling models. 19 00:01:35.145 --> 00:01:42.356 I'd soften that slightly — every lab rents spare capacity — but the pattern holds: the durable margin sits in owning megawatts and racks, and labs without a hyperscaler parent drift toward infrastructure to survive. Watch where the capex goes, not the model announcements. 20 00:01:42.356 --> 00:01:46.068 And one quiet infrastructure move: Arcee AI is moving all model and dataset storage off S3 onto Hugging Face in a multi-million dollar deal. 21 00:01:46.068 --> 00:01:50.548 Hugging Face converting distribution gravity into a storage business — every artifact that leaves S3 chips at AWS's grip on the ML data layer. Small deal, big precedent. 22 00:01:50.548 --> 00:01:55.161 On models: Gemini 3.5 Live Translate now serves real-time translated speech across more than seventy languages and two thousand language pairs, through AI Studio and the API. 23 00:01:55.161 --> 00:02:00.966 That replaces the ASR-to-MT-to-TTS pipelines teams stitched by hand for years. If you sell dedicated speech translation, your product just became an API parameter; if you consume it, your integration shrank to one call. 24 00:02:00.966 --> 00:02:06.852 Meanwhile Apple rebuilt Siri's architecture around Gemini under the hood, with a new Core AI framework for developers. And the EU Commission denied Apple's exemption request, so the new Siri won't ship in Europe at launch. 25 00:02:06.852 --> 00:02:15.255 The interesting admission is Apple choosing to rent frontier capability rather than build it — the practical questions for builders are what Core AI actually exposes and at what on-device latency. And if your roadmap assumed Siri integration for European users, you now have a geography-shaped hole with no fill date. 26 00:02:15.255 --> 00:02:20.664 Security: attackers compromised Microsoft's open-source tooling to harvest credentials specifically from AI developers — a supply-chain attack aimed at people holding API keys and training-cluster access. 27 00:02:20.664 --> 00:02:25.913 Targeting is the tell: API keys are compute, and compute is money. If you touched the affected packages, rotate now, and treat model-provider keys as tier-one secrets — attackers clearly already do. 28 00:02:25.913 --> 00:02:32.487 So the takeaway for anyone evaluating Fable 5 this week: score three axes in parallel — task accuracy against your default, refusal rate on your own domain prompts, and serving cost trajectory, because economics just moved one hundred x in a month. 29 00:02:32.487 --> 00:02:37.551 And write the numbers down. A model decision locked on benchmarks alone gets re-litigated within a month; locked on all three, you have something to check against when the next release lands. 30 00:02:37.551 --> 00:02:41.872 What I'm watching next: whether OpenAI or Google ships their own vetted tier this quarter — if they do, access vetting just became the industry's new pricing page. 31 00:02:41.872 --> 00:02:45.000 And watch your own refusal logs, because that's where you'll find out which side of the trust line you're actually on.