Anthropic retires Claude Opus 4.1, Claude Code fixes a Bash permission bypass, Groq pulls its public price list

Anthropic has retired claude-opus-4-1-20250805; every request to it now returns an error, with Claude Opus 5 as the recommended replacement. Services still pinning this model ID will not degrade silently, they fail outright. Researchers who need continued access can apply through the External Researcher Access Program.

Claude Code fixed a Bash permission bypass where a crafted command could hide parts of itself from permission checks. Teams that rely on permission modes to bound what an agent can run should upgrade now, especially for instances running unattended or in shared environments.

Supabase's logs.all Management API endpoint goes away on September 23; log querying moves to a new ClickHouse-backed endpoint that accepts ClickHouse SQL only. Scripts, integrations, and monitoring tools calling the endpoint directly need their queries converted before the cutoff. The announcement landed in late July and fell into the gap between two issue windows, so it takes its slot now, with six weeks left before the deadline.

Kimi K3 is now generally available in GitHub Copilot, hosted by GitHub on Fireworks AI and billed by usage: $3 per million input tokens, $15 output, $0.30 cached input, matching the Together AI rates from last issue. It ships off by default for Business and Enterprise plans, so administrators have to enable the policy before anyone in the organization can select it.

The groq.com/pricing page now carries the $650M fundraise announcement and positioning copy; the per-model token price tables are gone. Comparing costs now means digging through the console or docs, so teams with Groq in their cost model should keep their own copy of the last published rates.

OpenAI's Fast mode now takes long-context requests: prompts beyond 272K tokens on GPT-5.6 Sol, Terra, and Luna can run in the Fast tier, at up to 2.5x the speed of Standard. Workloads that fell back to Standard purely because of context length are worth re-benchmarking.

Zep added a 'Zep for Emerging Companies' tier: $13,000 for the first year, aimed at companies that have raised between $1M and $10M, with SOC 2 Type II, HIPAA BAA, DPA, guaranteed rate limits, and BYOK via AWS KMS included. Teams stuck between standard tiers that lack compliance features and an enterprise contract they cannot justify get a middle option.

LangSmith's Managed Deep Agents entered public beta: deep agents can now run hosted on LangSmith rather than on infrastructure you operate yourself. Teams already building long-running agents in the LangChain ecosystem can weigh the managed option against self-hosting.

Replicate added Nvidia H200 GPUs in single, 2x, 4x, and 8x configurations, with 144GB of memory per card at $0.001525/sec (about $5.49/hr) for a single GPU. One catch: H200 capacity requires a committed-spend contract, it is not on-demand.

The Gemini API released two embodied-reasoning endpoints in public preview, gemini-robotics-er-2-preview and a streaming variant, covering spatial reasoning, multi-robot coordination, and low-latency interaction. The same changelog carries an easy-to-miss deadline: gemini-robotics-er-1.6-preview shuts down on August 31, leaving projects on 1.6 the rest of this month to migrate.

langchain-anthropic fixed caller-supplied tool_choice not being preserved on its way to the API. Before the fix, code forcing the model toward a specific tool this way was not necessarily taking effect; after upgrading, behavior returns to what callers intended, so paths that rely on it deserve a regression pass.