Anthropic just flipped the switch back on its strongest models. Fable 5 and Mythos 5 are live again after the U.S. export control pause, now with tighter classifiers and credit-based access for many plans. The return happened July 1 with a gradual rollout through the week.

At the same time the company dropped Claude Sonnet 5, a new mid-tier model priced at $2 / $10 per million tokens. It targets exactly the workloads coding agents actually run: planning, tool use, long-horizon tasks, and research that previously required Opus-class models. Early reports show it closing much of the gap to Fable 5 at roughly 60% lower cost.

OpenAI meanwhile previewed its GPT-5.6 family on June 26 — Sol (flagship), Terra (balanced), and Luna (fast/cheap). The three models stay limited to roughly twenty government-vetted partners for now, with wider access expected mid-to-late July. Sol is already being tested on Cerebras wafer-scale hardware for extreme throughput.

The bigger story for people shipping code with agents isn’t just the raw models. It’s the harnesses around them. Multiple reports this week hammer home that production coding performance now depends heavily on workflow design, verification loops, sub-agent orchestration, and credit/session controls rather than picking the single “best” model.

GitHub’s July 1 changelog added agent-focused features to Copilot (session streaming preview, new Kimi K2.7 Code GA) while retiring GitHub Models on July 30. Platform billing and control surfaces are shifting fast.

China-side models continue squeezing margins. GLM-5.2 and recent Qwen releases keep delivering strong coding and agentic results at lower prices, though enterprise buyers still flag data governance concerns.

The practical takeaway: if you’re running coding agents daily, test Sonnet 5 on your harness first. The price/performance delta is real, Fable 5 is back for the hardest tasks, and the GPT-5.6 preview gives another data point on how the frontier is splitting into capability tiers. Harnesses that can route, verify, and budget across these options will pull ahead.

Primary sources: Anthropic Sonnet 5 announcement, OpenAI GPT-5.6 preview, recent coverage on Fable/Mythos redeployment and coding agent benchmarks.