Mythos 5 Partly Unblocked, Google's AI Talent Bleed, GPT-5.6 Gated

Good morning. It's Sunday, June 28, and we're covering the US government's partial reversal of the Anthropic model ban, the week Google lost four senior AI researchers to rivals, and OpenAI's GPT-5.6 launch that the government is still gating.

Alphabet's market cap shed roughly $270 billion over the week following those departures - the sharpest AI-talent repricing Wall Street has put on the board.

In today's brief:

• The US cleared Mythos 5 for critical-infrastructure teams two weeks after banning both Anthropic models - Fable 5 stays blocked
• Google DeepMind lost four senior researchers to OpenAI and Anthropic in six days, including a Nobel laureate and a Transformer co-author
• OpenAI previewed GPT-5.6 Sol on June 26 to government-vetted partners only - general availability still weeks out
• DeepSeek's open-source DSpark framework accelerates V4 inference 60-85%, and the full training stack is now public

THE BIG ONE

Anthropic's Mythos 5 Returns for 100+ Organizations - Fable 5 Remains Banned

Two weeks after the US Commerce Department ordered Anthropic to disable both Fable 5 and Mythos 5 for foreign nationals, the government on June 27 gave Anthropic a partial green light. Mythos 5 is now available to more than 100 US organizations that operate and defend critical infrastructure - "Annex A" entities in Commerce Department language - plus their foreign-national employees and Anthropic's own non-citizen staff. Fable 5 remains fully banned.

The ban began June 12 when Commerce Secretary Howard Lutnick sent a letter to Dario Amodei after a company claimed to have jailbroken Mythos, alarming the administration about national security risks. Anthropic shut down both models immediately, describing the action as a misunderstanding and beginning negotiations. Senior Anthropic engineers met Commerce officials face-to-face starting June 16. The partial lift followed on day 15 of the shutdown.

The practical impact: most enterprise teams at qualifying US critical-infrastructure companies can now access Mythos 5 again. International teams and consumer developers building on Fable 5 remain locked out with no restoration date announced. Our full coverage: why the US government pulled Fable 5, the initial suspension and what it meant for production systems, and the regulated-availability playbook for GovCloud teams.

Why it matters: The Annex A carveout is narrower than it sounds - most international dev teams and consumer builders are still blocked from Fable 5, and Mythos 5 access requires qualifying as critical infrastructure. If you were waiting for a full restoration, that has not come.

TALENT

Google Loses Four Senior AI Researchers in Six Days - Alphabet Sheds $270 Billion

The week ending June 27 was the worst stretch of talent loss in Google DeepMind's history by any public measure. Noam Shazeer - a co-author of "Attention Is All You Need," the 2017 Transformer paper behind nearly every frontier model - told Google on June 18 he was leaving for OpenAI. John Jumper, the 2024 Nobel laureate who led AlphaFold, departed for Anthropic on June 19 after nine years at the company. Two more Gemini researchers followed: Jonas Adler, Google's AI coding lead, and Alexander Pritzel, a pretraining specialist, both joined Anthropic on June 24.

Alphabet shares fell more than 5 percent on June 22 as analysts questioned whether Google can retain the people who build its best models. The combined market-cap loss across the week was roughly $270 billion. Gemini 3.5 Pro, expected before month-end, has been delayed to July.

The compensation dynamic is the driver: Anthropic and OpenAI are both approaching IPOs, and pre-IPO equity at companies valued in the hundreds of billions is a retention mechanism that even senior Google salaries struggle to match. Engineers at DeepMind have been moving to Anthropic at a ratio of roughly 11 to 1 in the last year.

Why it matters: Adler led Google's AI coding product work - that departure lands at exactly the moment the AI coding tool competition is most intense. Gemini 3.5 Pro's delay is a direct consequence, and the broader question of whether Google can win the coding-tool race without its coding-focused researchers is now live.

PLATFORMS

GPT-5.6 Sol Previews Under Government Supervision

OpenAI on June 26 launched GPT-5.6 - three models named Sol (flagship), Terra (balanced, 2x cheaper than GPT-5.5), and Luna (fast, lowest cost) - but limited rollout to roughly 20 pre-approved organizations at the explicit request of the US government. The June 2 executive order from President Trump asked frontier AI labs to voluntarily give federal agencies up to 30 days of pre-release access to models with advanced cyber capabilities; OpenAI is complying. CEO Sam Altman told staff the government is "approving access customer by customer."

Sol ships with OpenAI's most robust safety stack and introduces ultra mode - a subagent-backed reasoning layer for complex long-horizon work, similar in concept to Fable 5's extended thinking. General availability across ChatGPT, Codex, and the open API is expected in the coming weeks. The full changelog breakdown has dated context, and the June Codex vs Claude Code comparison weighs the two stacks as they stand now.

Why it matters: The dual-lab government preview - Mythos 5 and GPT-5.6 both launched under US oversight frameworks in the same two-week window - is now the established pattern for frontier model releases. Budget a 2-4 week government evaluation period into your planning for any new frontier model adoption.

RESEARCH

DeepSeek Opens DSpark: 60-85% Faster Inference for V4

DeepSeek released DSpark, a speculative decoding framework that accelerates per-user generation 60-85% on V4 Flash and 57-78% on V4 Pro. The HN thread pulled 762 points.

The approach is "semi-parallel" decoding: DSpark generates multiple candidate tokens in parallel, then selectively verifies the promising guesses instead of waiting for one-by-one autoregressive generation. Acceptance lengths improved 16-30% over previous methods including Eagle3 and DFlash. The DeepSpec codebase is now fully open-source and supports Qwen3 and Gemma model families for training and evaluation.

Why it matters: Speculative decoding is maturing from a research technique into the default inference optimization layer for production. Open-sourcing the training stack means you can apply DSpark's approach to your own fine-tunes, not just DeepSeek's hosted API.

TOOLS WORTH A LOOK

1. Vercel AI SDK 7 - Production agent runtime adding WorkflowAgent for durable execution, typed tool context, human-in-the-loop approvals, and telemetry integrations for Datadog, Langfuse, Braintrust, and Sentry. Open source. Our guide to the upgrade.
2. Perplexity Bumblebee - Read-only supply chain scanner for npm, PyPI, Go modules, and MCP configs. Checks against known-compromised packages without running install scripts. Apache 2.0, single Go binary. Our walkthrough.
3. Codex Remote GA - Now generally available on all ChatGPT plans. Start or continue work on a connected Mac or Windows host from iOS or Android, with one-to-one QR device pairing for authentication. Free tier included.

WHAT ELSE IS HAPPENING

• Exploitarium dump, 817 HN points: An anonymous GitHub account posted working exploits for a libssh/libssh2 heap overflow (CVE-2026-55200, CVSS 9.2) and a Gitea global admin bypass (CVE-2026-20896, CVSS 9.8) before advisories existed. The account was banned; the exploits are still circulating. Patch if you self-host Gitea or use libssh for CI/CD transport. (2 min read)
• Google caps Meta's Gemini access: Google told Meta around March it could not meet the capacity Meta sought to purchase. Meta instructed staff to use AI tokens more efficiently. The constraint is infrastructure, not policy - Google is enforcing similar limits with other large buyers.
• GPT-4.5 retired from ChatGPT: Retired June 27 after a 30-day sunset period. Existing conversations automatically migrated to GPT-5.5.
• Gemini 3.5 Pro pushed to July: Expected before June 30, now delayed. Google cited "roadmap reprioritization" without specifics.
• Simon Willison: 6,000 prompt injections, zero successes: A public experiment attracted 6,000 injection attempts against an Opus 4.6 instance. None extracted the secret. His takeaway: frontier model anti-injection training has improved meaningfully, but the attack surface grows with every new tool an agent can call. (10 min read)

FROM THE SITE

What We Published Today

The Vercel AI SDK 7 production agent guide covers everything new: WorkflowAgent for durable execution that survives process crashes, typed tool and runtime context, human-in-the-loop approvals, telemetry integrations, realtime voice, and the migration path from AI SDK 6.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.