Fable 5 Comes Back Online, Hermes Learns to Judge Itself, and Short-Form Video Gets a Director.
Eighteen days after the US government cut it off, Claude Fable 5 is back — alongside a new Sonnet 5 that undercuts it on price. OpenAI answers with a government-gated GPT-5.6 family. Hermes Agent ships Mixture-of-Agents as a real model you can pick. Manus goes fully independent and turns toward small business.
The last round ended on a question: could "multi-model, multi-jurisdiction" survive contact with reality? This week answered it. Claude Fable 5 came back online exactly as the policy environment predicted it might — restored, but with new usage rules, new safeguards, and a new sibling model designed to make the whole conversation about a single frontier model less urgent.
Underneath the policy story, the tooling kept moving. Hermes Agent shipped a release that changes how multi-model ensembles work in production. Manus, freshly cut loose from Meta, pivoted hard toward the small-business segment it always quietly served best. Here's the full read on the week of June 26 – July 2, and what to do about each piece.
Fable 5 and Mythos 5 return. Sonnet 5 launches the same day.
Eighteen days after the US Department of Commerce forced Anthropic to disable Fable 5 and Mythos 5 worldwide, both models are back. Commerce Secretary Howard Lutnick informed Anthropic on June 30 that export controls were lifted, citing the company's cooperation on protocols, standards, and future releases. Fable 5 returned globally on July 1 across claude.ai, Claude Platform, Claude Code, and Claude Cowork.
How we got here
Anthropic's account of the incident, backed by testing it says it ran with Amazon, Microsoft, and Google: multiple less-capable models — including Opus 4.8, GPT-5.5, and Kimi K2.7 — could identify the same vulnerabilities, and every tested model could reproduce the exploit demonstration. Anthropic is now proposing an industry-wide framework for scoring jailbreak severity.
What changed operationally
Claude Sonnet 5: the model built to make this less fraught
Launched the same day, Sonnet 5 is now the default model for Free and Pro plans. Pricing: $2/M input and $10/M output through August 31, rising to $3/$15 after — roughly 40–60% cheaper than Opus 4.8. Anthropic describes it as "the most agentic Sonnet model yet."
| Model | Input / M | Output / M | Positioning |
|---|---|---|---|
| Sonnet 5 | $2.00* | $10.00* | Agentic default · *intro price to Aug 31 |
| Opus 4.8 | $5.00 | $25.00 | Higher accuracy, harder tasks |
| Fable 5 | $10.00 | $50.00 | Frontier ceiling, safeguarded |
OpenAI answers with GPT-5.6 — Sol, Terra, Luna — behind a government-coordinated gate.
Four days before Fable 5's restoration, OpenAI previewed GPT-5.6 — not one model but three durable tiers: Sol (flagship), Terra (balanced, ~half GPT-5.5's cost), and Luna (fast, high-volume). New: a max reasoning effort setting, and an ultra mode that coordinates subagents to divide and accelerate complex work.
Unlike a normal launch, GPT-5.6 shipped to roughly 20 trusted partners via API and Codex only, coordinated with the US government. OpenAI says the models are classified High capability under its Preparedness Framework for both cybersecurity and biological/chemical risk — a first — though below the "Critical" threshold. On Terminal-Bench 2.1, Sol and its ultra mode reportedly surpass Claude Mythos 5; Terra surpasses Fable 5; Luna surpasses Opus 4.8 — though these are OpenAI's own preview figures, not independently audited.
OpenAI plans to bring Sol, Terra, and Luna to general availability "in the coming weeks," with Sol on Cerebras hardware in July at up to 750 tokens/second. For now, this is the frontier's most capable model family sitting in a waiting room — a mirror image of what just happened to Fable 5.
| Tier | Input / M | Output / M | Positioning |
|---|---|---|---|
| Sol | $5.00 | $30.00 | Flagship · same price as GPT-5.5 |
| Terra | $2.50 | $15.00 | Balanced · ~half GPT-5.5 cost |
| Luna | $1.00 | $6.00 | High-volume · fast, cheap |
Hermes Agent v0.18.0 — "The Judgement Release."
Nous Research shipped v0.18.0 after a week-and-a-half priority sprint that closed every open P0 and P1 across the repo. On top of that stability sweep, the release is genuinely about how well Hermes reasons and how it knows when work is actually finished — the "judgement" in the name.
Mixture-of-Agents landed as MoA 2.0 in late June and already showed real lift: a default preset blending GPT-5.5, DeepSeek, and Opus scored roughly 8% higher than Opus alone and 11% higher than GPT-5.5 alone. What v0.18.0 changes is packaging — MoA becomes a model you select once and forget.
What's new
Manus goes independent — and leans all the way into small business.
China's regulators blocked Meta's proposed $2B acquisition of Manus in April 2026; by June 11, Meta had fully unwound the relationship. Meanwhile the underlying business kept growing: Manus hit an estimated $450M in annualized revenue in June 2026, up from $127M a year earlier.
Against that backdrop, "Manus for Small Business" reads less like a new feature and more like a strategic homecoming. Manus's most-cited real-world success stories were never enterprise — a Singapore florist used it to build an online storefront, write product descriptions, manage inventory, and produce marketing materials from scratch.
What the SMB push likely means in practice
- Storefront + inventory + marketing in one agent run — the florist pattern productized.
- Business-plan and go-to-market generation — tuned to local demographics and loan-application-ready structure.
- Lower-friction pricing tiers — expect positioning against Starter/Standard tiers.
- Distribution through Manus Academy — the existing free e-learning platform targets small-business owners specifically.
Shorts Studio: an AI editor built to make one clip work everywhere.
Higgsfield's pitch: "Adapt any clip to the formats people watch, with pacing and editing built to grab attention from the first second." Shorts Studio is powered by Gemini Omni Flash — Google's any-to-any multimodal model, which reasons across text, image, audio, and video in a single prompt. It works best with real footage as input — the point is re-cutting and re-pacing what you already shot.
What makes Omni Flash different from a normal video editor
- Conversational multi-turn editing — iterate "change the angle," then "warm the lighting," as a running conversation.
- Grounded in real-world knowledge — physics, gravity, and momentum render correctly rather than plausibly.
- Native multimodal input — feed it a clip, a reference image, and a text note in one prompt.
- SynthID watermarking — every output carries Google's invisible AI-provenance watermark automatically.
The benchmark math just got harder to trust at a glance.
Worth a note this round, since two of the week's biggest claims (GPT-5.6's Terminal-Bench wins, Fable 5's frontier reclaim) both lean on benchmark numbers that are becoming genuinely difficult to compare across labs.
Scale AI's SWE-bench Pro — 1,865 real-world software tasks — is meant to resist contamination. But as of late June, three different numbers all claim to be "the" SWE-bench Pro score: 59.1% (GPT-5.4 xHigh), 69.2% (Claude Opus 4.8, vendor-reported), and 47.1% (Claude Opus 4.6, Scale's private set). The gap isn't models getting worse — it's different scaffolding producing different, all "correct," answers.
What Fable 5 hit before the suspension
Fable 5 briefly held all four marks before the 18-day suspension took it off the board entirely. Those scores are now live again as of July 1.