Anthropic Launches Sonnet 5 and Restores Fable 5 as Multi-Company Jailbreak Framework Proposed

Anthropic launched Claude Sonnet 5 on June 30 ¹, pricing it at $2 per million input tokens and $10 per million output tokens through August 31, after which prices rise to $3 and $15. The model replaces Sonnet 4.6 as the default across Free and Pro plans on claude.ai, Claude Code, and the Claude Platform. On OSWorld-VerifiedA benchmark measuring autonomous desktop and browser control by AI agents, including multi-step task completion and checkout flows. (a benchmark measuring autonomous desktop and browser control), Sonnet 5 posts 81.2%, up from Sonnet 4.6’s 78.5%; on SWE-bench Pro (agentic coding), it reaches 63.2% against Sonnet 4.6’s 58.1% and Opus 4.8’s 69.2%; on GDPval-AAArtificial Analysis's knowledge-work benchmark, scoring models on realistic multi-step professional output tasks. (knowledge work), it scores 1,618 — narrowly above Opus 4.8’s 1,615. Artificial Analysis, which evaluated the model prior to release, found that Sonnet 5 uses approximately 40% more output tokens per task than Sonnet 4.6 and three times the agentic turns on knowledge-work evaluations; at introductory $2/$10 pricing, cost per task runs below Opus 4.8, while at standard $3/$15 pricing it runs approximately 15% higher. Testers at Zapier, Lovable, Kiro, and Pace (an insurance-workflow automation firm) reported that Sonnet 5 completed multi-step agentic tasks — including a Salesforce update combined with enterprise email dispatch in a single autonomous run — where Sonnet 4.6 had not finished.

Sonnet 5 arrives one week after Gemini 3.5 Flash gained computer use and browser control at approximately 30% of GPT-5.5’s cost (2026-w26), and four days after OpenAI previewed GPT-5.6 Sol with restricted access and stronger cyber safeguards. The three releases complete a period in which all three major lab families published a cost-optimized agentic model. Sonnet 5 improves on Sonnet 4.6 across every published benchmark and shows lower hallucination, sycophancy, and prompt-injection susceptibility rates than its predecessor — properties directly relevant to agentic commerce deployments where models interact with untrusted third-party web content and merchant interfaces. Anthropic notes that Fable 5 and Mythos 5 retain higher accuracy on the most demanding tasks; at standard pricing, Sonnet 5 costs less per token than GPT-5.5 and Gemini 3.1 Pro but more than Gemini 3.5 Flash.

US export controls on Fable 5 and Mythos 5, applied June 12 (2026-w24) after Amazon researchers reported a classifier bypass technique, were lifted June 30 ². The US Department of Commerce’s Center for AI Standards and Innovation (CAISICenter for AI Standards and Innovation — a unit of the US Department of Commerce that assessed Anthropic's updated safety classifiers for Fable 5.) reviewed Anthropic’s updated classifiers over two weeks; the revised classifier blocks the reported bypass in over 99% of cases. Commerce Secretary Howard Lutnick signed the reversal; Anthropic committed to pre-release government access, rapid information sharing, and joint research with designated agencies. Fable 5 returned to global access July 1 at up to 50% of weekly usage limits through July 7, transitioning to usage credits thereafter. That same day, Anthropic, Amazon, Microsoft, and Google proposed a four-criterion severity framework for AI jailbreaks ³: capability gain (how far a jailbreak extends an attacker’s access beyond available tools), breadth of that gain, ease of weaponization, and discoverability. Anthropic opened a HackerOne program for Fable 5 vulnerability reporting. The framework formalizes pre-release government testing and interagency vulnerability sharing — no comparable multi-company governance mechanism had previously been proposed at this scope.

Mastercard’s Singapore ‘Next Lap in Payments’ Innovation Circuit opened July 2 ⁴, the fourth edition of the company’s primary Asia Pacific co-creation venue for agentic commerce infrastructure. The Singapore Experience Center — one of seven globally — brings together financial institutions, payment partners, merchants, and policymakers on agentic commerce, trusted digital identity, tokenisation, interoperability, and AI-driven network intelligence. Mastercard confirmed plans for centres in Kuala Lumpur and Tokyo; those markets join Singapore, Malaysia, and Thailand, where live Agent Pay transactions completed on March 4 (DBS, UOB, CIMB, RHB) and April 7 (Krungthai Card) respectively. The three model launches of the past two weeks — Gemini 3.5 Flash with computer use (2026-w26), GPT-5.6 Sol, and Sonnet 5 ¹ — each extend browser-based agentic execution across price tiers; the Fable 5 restoration ², the jailbreak severity framework ³, and the Singapore Innovation Circuit ⁴ mark the governance, security, and infrastructure layers advancing in parallel. No jurisdiction has published regulation specific to agent-initiated commercial transactions.