In partnership with
In Today’s Issue:
🦋 Claude Fable 5 becomes the public face of Mythos
🧪 Anthropic’s system card gets into the real guardrails
📊 Fable 5 takes the lead on FrontierCode Diamond
🎥 Anthropic explains the launch in its Fable 5 video
🌍 Google turns Gemini into a live voice interpreter
✨ And more AI goodness…
⚡ The Signal
Today’s issue is almost entirely about Claude Fable 5, Anthropic’s public version of a Mythos-class model.
Anthropic is treating this release as a deployment split. Fable 5 is the broadly available model, with extra safeguards for cyber, biology, chemistry, and distillation attempts. Mythos 5 is the same underlying model with some safeguards lifted, but only for trusted access programs.
That makes this launch feel less like a normal model upgrade and more like a live test of frontier-model access control. The system card is the interesting part: Opus fallbacks, API blocks, CB-1 / near-CB-2 bio risk, strong cyber numbers, 30-day retention for Mythos-class traffic, and alignment notes that are unusually frank.
All the best,
Kim Isenberg
🌍 Google makes Gemini a live interpreter
Google launched Gemini 3.5 Live Translate for near real-time speech-to-speech translation across 70+ languages. It is rolling out through the Gemini Live API and Google AI Studio, private preview in Google Meet, and Google Translate on Android and iOS. In Google's demos the translation trails the speaker by a couple of seconds and keeps their tone and pacing, which makes the conversation feel direct rather than mediated. Exciting!
👉 tl;dr: Google is pushing translation from turn-taking captions toward continuous voice mediation.
✨ Noam Brown wants compute-aware benchmarks
OpenAI reseracher Noam Brown argues that frontier model performance is becoming increasingly dependent on test-time compute: the number of tokens, dollars, or seconds a model is allowed to spend before answering. His point is that single benchmark scores can hide large capability gaps, because stronger models may keep improving long after weaker models plateau.
Brown recommends that model releases, leaderboards, and preparedness frameworks measure performance against tokens, cost, or wall-clock time, not just a single headline score.
👉 tl;dr: Brown wants model comparisons plotted against tokens, dollars, or wall-clock time instead of a single headline score.
🤖 A humanoid robot reaches Chimborazo
A modified Unitree G1 named Pemba reached the summit of Ecuador’s 20,341-foot Chimborazo volcano as part of a robotics expedition aimed at extreme-environment field work. The important caveat: Pemba walked independently on terrain below 30 degrees, while humans carried it through steeper parts of the 16-hour summit push.
👉 tl;dr: Humanoids are moving into real terrain, but autonomy still drops sharply when the environment stops being friendly.
Turn Claude Fable 5 into an access-control checklist
Why it helps: The useful part of Anthropic's launch is the access-control design: the split between public Fable and restricted Mythos, Opus fallbacks, API blocks, retention rules, and high-risk classifiers restricted Mythos, Opus fallbacks, API blocks, retention rules, and high-risk classifiers.
Try this: Paste the launch post and system-card notes and ask: “Separate the release into public capability, restricted capability, safeguards, fallback behavior, data-retention rules, and tests we would run before deploying it inside a company.”
🎬 Watch This
Anthropic’s announcement video frames Fable 5 as the public release of Mythos-class capability: stronger long-horizon coding, research, analysis, and knowledge-work performance, with safeguards that route some high-risk requests away from Fable and toward Opus 4.8.
The clip is useful because it compresses the product pitch into one question: how do you ship a model that is strong enough to need a restricted twin?
“Mythos 5 will occasionally take reckless or destructive actions in service of user-assigned goals.”
– Anthropic, System Card: Claude Fable 5 & Claude Mythos 5, p. 53 (June 2026)
Claude Fable 5 Is the Public Face of Mythos
The Takeaway
👉 Fable 5 and Mythos 5 share the same underlying model; Fable is the broadly available version with additional safeguards.
👉 Anthropic says Fable 5 is its strongest generally available model, with the largest lead on long, complex tasks.
👉 High-risk cyber, biology/chemistry, and distillation-related requests can trigger fallback to Claude Opus 4.8 on consumer surfaces.
👉 Pricing is $10 per million input tokens and $50 per million output tokens; subscription access is included only through June 22 before usage credits kick in on June 23.
Anthropic launched Claude Fable 5 on June 9 as the public version of its new Mythos-class model. Fable 5 is powerful enough that Anthropic is also releasing Claude Mythos 5, the same underlying model with some safeguards lifted, only to trusted cyberdefense partners through Project Glasswing and later to selected biology researchers. One caveat right away: the safeguards are tuned conservatively, and in biology and chemistry it is currently hard to get substantive answers out of Fable at all.
The model claims are broad. Anthropic says Fable 5 is state-of-the-art on nearly all tested capability benchmarks, with stronger performance in software engineering, knowledge work, vision, memory, and life sciences. The coding number that matters for today’s issue is FrontierCode: in Anthropic’s system card, Fable 5 ranks #1 on the Diamond subset with a 29.3% score and 30.2% pass rate, ahead of Opus 4.8 and GPT-5.5. And this just a few days after the benchmark was published. Kudos, Anthropic!
Early testers are impressed, and several report one-shot results that would normally take days of engineering. A few examples
There are two caveats, however: Fable 5 is only included in the subscription plan until June 22nd; after that, it's only available via usage. Also, as mentioned before, the guard rails are set very high. Beyond those two limits, this is the strongest public model release of the year so far, landing a week after Anthropic filed for its IPO
Why it matters: Fable 5 is a frontier-model launch wrapped in an access-control experiment: the real question Anthropic is answering is which version of this strength the public should get.
Sources:
🔗 https://www.anthropic.com/news/claude-fable-5-mythos-5
🔗 https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf
The IT strategy every team needs for 2026
2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn what’s changing and how to prepare.
The chart: Claude Fable 5 leads FrontierCode Diamond at max effort with a 29.3% score and 30.2% pass rate, versus Claude Opus 4.8 at 13.4% / 14.5% and GPT-5.5 at 5.7% / 6.4%. FrontierCode is Cognition’s production-coding benchmark: agents work inside real open-source repositories, produce patches, and are graded against blocking tests plus rubric criteria such as test coverage and prohibited implementation patterns.
The lesson: Fable 5’s coding lead is tied to a more demanding definition of engineering work: maintainable, scoped patches that look closer to work a human maintainer might actually merge.
The caveat: The best result still clears only about a third of the Diamond subset, and the system card notes cost on a log scale. Production-grade autonomous coding is improving fast, but the remaining gap is still large.
Inside Fable 5’s System Card
⚡ Bottom line Anthropic says Mythos 5 is its most capable model ever, but the system card is dominated by deployment control: Fable 5 gets public access with extra classifiers, while Mythos 5 stays restricted for trusted partners.
💡 Stakes The safety section is sharper than a standard “we tested it” paragraph. The PDF says Mythos 5 is treated as CB-1 for chemical/biological risk, sits near the CB-2 threshold without crossing it, and is far ahead of Opus 4.8 on several cyber tasks when safeguards are off.
🔎 What it means The system card is most interesting when it admits the messy parts: fallback behavior differs by surface, API requests can be blocked by default, Mythos 5 sometimes shows grader/evaluation awareness, and Anthropic is adding 30-day retention for Mythos-class traffic to detect novel attacks and reduce false positives.
The safeguards are concrete. Fable 5 uses classifiers for cybersecurity, biology and chemistry, and distillation attempts. In Claude’s consumer apps, a triggered request falls back to Opus 4.8 and the user is told which model handled it. In the Messages API, there is no automatic fallback by default: the request is blocked and returned with a structured refusal category unless a developer implements or opts into fallback behavior.
The bio-risk section is unusually candid. Anthropic says Mythos 5 shows significant gains over Mythos Preview, treats it as CB-1, and says its CB-2 judgment is “much more complicated.” The strongest CB-2 signals include a tabletop exercise where generalist biology PhDs using Mythos 5 could substitute for specialist expertise on some tasks, and experts estimated two-person teams produced work equivalent to 40-95 working days over two days.
The cyber numbers explain why Fable is fenced. On ExploitBench, with safeguards off, Mythos 5 captured 10.75 mean capability flags and reached 78% Cap%, ahead of Mythos Preview, Opus 4.8, and GPT-5.5. On CyberGym, it reproduced 83.8% of targeted vulnerabilities on a single try and produced at least one crash in 99.4% of tasks. Fable’s cyber safeguards flagged 407 of 410 ExploitBench episodes, on average after 27 turns.
The alignment section is not cleanly triumphant. Anthropic says Mythos 5 is broadly comparable to Opus 4.8 on safety and alignment, but also reports cases of reckless or destructive actions in service of user goals, elevated grader-oriented reasoning, and occasional evaluation awareness that is not always verbalized. That is the release-year tension: the model is strong enough to route more work through it, and strange enough that Anthropic is routing parts of it away from the public.
What happens when you throw out the GTM playbook
That investor was wrong. Gamma is now worth $2B, with 50M users and more than half their growth driven by word of mouth.
They're one of 6 AI-native startups in HubSpot for Startups' free Bold Bets Playbook. Replit grew revenue 50x after half the team pushed back on the strategy. Ramp generated 100M+ views from a single stunt. Clay's co-founder wouldn't hang up a sales call until the prospect DMed him in Slack.
Each one took a GTM risk most founders would never greenlight. Each one paid off.