Anthropic’s Fable Relaunch and Claude Sonnet 5 Shake Up AI Landscape

Fable is back on schedule. Anthropic relaunched Claude Fable 5 on July 1 after an 18-day global ban. The relaunch hit tooling fast—Cursor reports Fable 5 leads its evaluations.

Devin deployed Fable 5 across cloud, desktop, and command line interfaces. Perplexity restored it as an orchestrator model. Anthropic reset rate limits for all users. “Claude Fable 5 is back,” the company declared.

Builders are shifting toward multi-model orchestration. Fable 5 handles high-value reasoning and planning while delegating tasks like implementation and verification to other models. This division improves efficiency and model specialization.

Fable 5 scored 16.10% on the Remote Labor Index. Meanwhile, Sonnet 5 launched June 30 with a 1 million token context window and agentic coding accuracy of 63.2%. It approaches Opus 4.8’s 69.2% performance but at 60% lower cost. Pricing is $2 per million tokens input and $10 per million tokens output, down from Sonnet 4.6’s higher rates.

Sonnet 5’s tokenizer can expand input tokens by up to 35%. It also features improved safety with lower hallucination, reduced sycophancy, and better refusal of malicious requests. California recognized this value, signing a deal with Anthropic on June 29 offering 50% off Claude for all state agencies. The deal includes free workforce training and developer support.

California deployed Claude at DMV, Department of Healthcare Services, and other agencies. The state AI assistant “Poppy” is piloted by over 2,800 employees across 67 departments. Governor Gavin Newsom highlighted this partnership as a milestone for government AI adoption.

Anthropic’s Claude Science workbench entered beta on June 30. It integrates 60+ databases and scientific tools for research applications. Anthropic also launched a drug discovery program targeting neglected diseases, leveraging AI for social impact.

OpenAI quietly released GPT-5.6 in a limited preview controlled by the US government. Its three models—Sol, Terra, and Luna—offer different price points and capabilities. Sol Ultra scored 91.9% on Terminal-Bench 2.1, signaling strong coding and reasoning skills.

Meanwhile, GLM-5.2 leads open models with 55.3% Pass@1 on APEX-SWE Integration and is building an ecosystem including ZCode, a cross-platform dev environment with BYOK support and quota boosts for subscribers. LangChain published guides to use GLM-5.2 in coding workflows, and developers are adopting it as a daily driver.

Hardware and decoding improvements keep pace. The vLLM project added native DSpark speculative decoding for DeepSeek models, hitting roughly 250 tokens per second on 8×B300 GPUs. mgoin_ previewed GLM-5.2 DSpark with 1.5× faster decode. Qwen3-32B showed 50% higher throughput on the same hardware.

MolmoMotion, an open-source model predicting 3D object movement from video, outperformed peers on the PointMotionBench. It includes model weights, training code, and a dataset of 1.16 million videos, reflecting open science progress in multimodal AI.

VirBench research revealed a stark truth: AI accuracy on viral sequence retrieval jumped from 16.9% to 92.8% by adding a deterministic retrieval tool, not by changing models. Anthropic published this finding, underscoring the power of structured data access to improve AI results.

Finally, the federal government continues to cast a wary eye on Anthropic, designating it a “supply-chain risk” after Pentagon contract refusals. Yet export controls on Claude Fable 5 and Mythos 5 lifted June 30, allowing global relaunch. Amazon researchers had found jailbreaks that triggered these controls. Anthropic responded with a classifier blocking over 99% of jailbreak attempts.

The AI scene is busy. Relauched models, new benchmarks, government deals, and open research all push the field forward. Fable 5’s comeback and Claude Sonnet 5’s pricing and performance shake up market dynamics. Meanwhile, structured data and orchestration strategies reshape how AI works in practice.