WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-06-12 1 00:00:00.000 --> 00:00:10.240 Anthropic's Fable 5 just posted the best coding numbers ever recorded — and got caught silently swapping in a weaker model when it didn't like your question. 2 00:00:10.240 --> 00:00:27.920 It's Friday, June 12, 2026. Here's the rundown: Anthropic moves to own its own servers, agent-skill scanners turn out to barely agree with each other, an Oracle flaw breaches a hundred-plus companies, and Ideogram ships a top-ten image model as open weights. 3 00:00:27.920 --> 00:00:51.680 Fable 5 first. Eighty point three percent on SWE-Bench Pro against Opus 4.8's sixty-nine point two and GPT-5.5's fifty-eight point six. On Cognition's FrontierCode it more than doubles Opus — twenty-nine point three versus thirteen point four. 4 00:00:51.680 --> 00:01:05.680 And none of that is the story. The story is what happens when a classifier flags your prompt: Fable doesn't refuse, it silently routes the question to Opus 4.8 and hands you a weaker answer with no disclosure. 5 00:01:05.680 --> 00:01:25.280 The triggers were aggressive. IBM X-Force's Valentina Palmiotti says it rejects anything tangentially cyber related, including reading a blog post. An immunologist at Jackson Laboratory found the word cancer tripped the biosecurity classifier. 6 00:01:25.280 --> 00:01:37.840 The mechanism explains the mess. Fable 5 is the same underlying weights as Mythos 5 — the difference is classifier routing bolted on top, with the unrestricted Mythos tier reserved for Glasswing partners. 7 00:01:37.840 --> 00:01:58.000 Anthropic says safeguards fire in under five percent of sessions, tuned conservatively. And the under-covered admission in our briefing: they deliberately degrade answers on questions that might relate to AI development, so competitors can't use Fable for their own research. 8 00:01:58.000 --> 00:02:08.480 That's a competitive moat wearing a safety vest. To their credit, the retraction was fast — under two days from launch to making the safeguards visible instead of silent, after Wired broke it. 9 00:02:08.480 --> 00:02:23.120 Pricing is ten dollars per million input tokens, fifty per million output — double Opus. But it's included free on Pro, Max, Team, and Enterprise plans until June 22. 10 00:02:23.120 --> 00:02:43.520 So the move this week is obvious: run your hardest agentic workloads against it while the meter's off. But read the data terms first — Fable requires thirty-day retention on all traffic, even for enterprises that previously negotiated zero retention. For some compliance regimes that disqualifies it outright before you ever see a benchmark. 11 00:02:43.520 --> 00:02:54.480 On benchmarks — Endor Labs ran an independent harness and scored Fable mid-tier on coding, flatly contradicting Anthropic's first-party numbers. 12 00:02:54.480 --> 00:03:10.080 When the vendor's evals and a third party's disagree by that much, neither is your answer. Run your own. And the precedent here outlasts the model: a frontier lab returned a different model's answers without disclosure, and stopped only because researchers caught it. 13 00:03:10.080 --> 00:03:18.800 Meaning silent model substitution on hosted APIs is now documented practice, not a paranoid hypothetical. 14 00:03:18.800 --> 00:03:32.160 So expect routing-detection probes in eval suites, and no-silent-fallback as a procurement line item by Q4. Add a model-fingerprint check to your pipeline now — it's an afternoon of work and it just earned its keep. 15 00:03:32.160 --> 00:03:46.480 The story everyone will misread today sits right next to that one: The Information reports Anthropic is moving to control its own AI servers, attacking compute — its single largest expense. 16 00:03:46.480 --> 00:04:07.360 Read it against Fable's fifty-dollar output pricing. A lab that owns its serving stack can cut API prices without torching margin — and its cloud patrons, who are also its investors, lose a captive customer. Every frontier lab is converging on the same answer because rent on rented compute is the biggest number on the P&L. 17 00:04:07.360 --> 00:04:23.040 Downstream of that: Amkor breaks ground on a six-hundred-fifty-million-dollar phase one in Gwangju — the first of six packaging plants in Korea running through 2035, driven by TSMC order overflow. 18 00:04:23.040 --> 00:04:48.720 Advanced packaging, not wafer starts, has been the binding constraint on accelerator supply since CoWoS sold out. Real OSAT capacity outside Taiwan eases the bottleneck and chips at the single-point-of-failure problem — and a 2035 horizon means the packaging industry is underwriting a full decade of accelerator demand. That's the strongest demand signal on today's tape. 19 00:04:48.720 --> 00:05:11.200 Developer tools. NVIDIA shipped SkillSpector, an Apache-2 scanner that checks agent skills against sixty-four vulnerability patterns across sixteen categories — prompt injection, exfiltration, MCP tool poisoning. It gates NVIDIA's own Verified Skills catalog. 20 00:05:11.200 --> 00:05:22.320 The category is justified — cited research puts twenty-six point one percent of skills vulnerable and five point two percent likely malicious. The tool, less so. 21 00:05:22.320 --> 00:05:40.800 The OpenClaw dataset — sixty-seven thousand rows — has SkillSpector flagging forty-eight point seven percent positive while catching only six point eight percent of confirmed-malicious rows. VirusTotal caught seventy-two point eight. 22 00:05:40.800 --> 00:06:01.120 And no scanner pair agrees on more than ten point four percent of flags. So run it in CI, never as your only gate — two independent scanners, minimum. Same beat from a different angle: LWN documents an AI agent filing low-quality contributions across Fedora faster than humans can review them. 23 00:06:01.120 --> 00:06:09.520 The pattern holding across both: the agent supply chain generates work faster than tools or maintainers can vet it. 24 00:06:09.520 --> 00:06:16.960 If you maintain an open source project, write contribution rate limits and provenance requirements this month — not after your incident. 25 00:06:16.960 --> 00:06:31.200 Security, and this one's a clock, not a story. Google says a cybercrime gang exploited an Oracle flaw at scale and breached more than a hundred companies — victims have been notified. 26 00:06:31.200 --> 00:06:44.320 MOVEit and Citrix Bleed kept claiming victims for months after disclosure, and almost all of them knew and deferred. If you run the affected Oracle software, patch today — not this sprint, today. 27 00:06:44.320 --> 00:07:01.040 Launches. Ideogram 4.0 ships as open weights and debuts at number eight on the text-to-image leaderboard — a top-ten image model you can self-host with zero per-call fees. 28 00:07:01.040 --> 00:07:18.960 For anyone generating images at volume, the arithmetic just flipped toward a GPU bill over an API bill. Closed image APIs are about to get squeezed from below exactly the way open LLMs squeezed text pricing through 2025 — benchmark it against your current API this week. 29 00:07:18.960 --> 00:07:26.240 Quick hits. Solar generated more US electricity than coal for the first time on record. 30 00:07:26.240 --> 00:07:32.480 Homebrew 6.0.0 lands — first major release of the package manager since 2023. 31 00:07:32.480 --> 00:07:38.560 Xiaomi releases MiMo Code as open source, joining the open coding-model field. 32 00:07:38.560 --> 00:07:46.000 The macOS 27 beta breaks booting Asahi Linux on Apple silicon — hold the update if that's your daily driver. 33 00:07:46.000 --> 00:07:54.720 And a wargame study finds LLMs reach for tactical nukes in ninety-five percent of simulated conflicts. 34 00:07:54.720 --> 00:07:59.440 Keep them off the launch keys and out of your incident response runbooks, in that order. 35 00:07:59.440 --> 00:08:19.440 Our call: Anthropic cuts Fable 5's list price at least forty percent — output under thirty dollars per million tokens — by September 30, 2026, because the server buildout and the shared Mythos weights say ten-and-fifty is a placeholder, not a position. 36 00:08:19.440 --> 00:08:29.920 What proves us wrong: Anthropic's public pricing page still showing ten dollars in, fifty out on September 30 — it's in The Book, and it settles that day.