AI News Briefing — Anthropic releases Claude Haiku 5.5 at a tenth of the price

Model releases

[2026-10-07] Anthropic released Claude Haiku 5.5 (claude-haiku-5-5) on its API, Bedrock, Vertex AI and Azure. Prompts up to 100K tokens cost $0.10 per million input and $0.50 per million output, a tenth of Haiku 4.5’s $1/$5; past 100K the whole request costs $0.50 and $2.50. Cache reads are $0.01, or $0.05 above 100K. Anthropic reports 39.2% on Terminal-Bench 4.0 against GPT-6 Luna’s 16.4%. It is the first Haiku with an effort setting, and reasoning cannot be turned off. Simon Willison measured its new tokenizer at about 1.25 times as many tokens as Haiku 4.5 for the same prompt. (official, source)

Anthropic ran the Terminal-Bench comparison itself. Treat the gap to GPT-6 Luna as a claim until someone else reproduces it.

For Engineering Managers / Tech Leads: A request that crosses 100K tokens is billed at the higher rate in full, five times the short-prompt price. Re-price a week of real Haiku 4.5 traffic with the new tokenizer and always-on reasoning before booking the tenfold saving, and look hard at jobs that sit just above the line.
[2026-10-06] OpenAI’s Decisions API (POST /v1/decisions) is now in public beta, a week after its limited DevDay preview, with general availability promised “in the coming weeks.” It runs only on gpt-6-luna, returns typed predicate, choice or score answers with probabilities, and bills input alone at $0.10 per million tokens, with no output or cache charges. That price now sits beside the free open-weight rivals from Cloudflare and AWS. (official, source)

Routers and classifiers fit this shape best, since what they return is a label and a probability rather than prose.
[2026-10-07] OpenAI put GPT-6 into ChatGPT with “Intelligent UI”, answers that include interactive charts, calculators and buttons; Plus, Pro, Business and Enterprise get it first, Free and Go a day later. Users can turn the number of visuals down. No API change was announced. (official, source)

Coding agents

[2026-10-07] Local sandboxing for GitHub Copilot is generally available at no extra charge in Copilot CLI, the Copilot app and VS Code Agent Host sessions, on Windows, macOS and Linux. Policies limit which paths agent commands can touch, network access and Git credentials, and can cover local MCP servers; enterprises can require it. Copilot CLI 1.0.94 also lists models from a running Ollama instance in /model, though offline mode still needs COPILOT_OFFLINE=true. (official, official)

Path, network and credential limits now ship inside Copilot on all three desktop platforms, with no separate container setup.

For Security Engineers: Enterprises can now require the sandbox rather than hope each developer turns it on. A sensible first policy blocks network access and Git credentials for agent commands, with exceptions added only for the local MCP servers the team actually runs.
[2026-10-07] GitHub switched AI-detected secret scanning to a purpose-built model that reads surrounding code to catch credentials with no recognizable token format, such as plain passwords. It is included with Secret Protection and Advanced Security; AI push protection is in preview and will draw on GitHub AI Credits once enabled. GitHub gave no precision or recall figures. (official)

With no precision or recall published, each team will learn the false-positive rate on plain-password findings from its own alert queue.
[2026-10-07] Claude Haiku 5.5 is generally available in GitHub Copilot on paid plans, billed at provider list price. (official)

MCP

[2026-10-07] Google’s MCP Toolbox Java SDK reached 1.0 (com.google.cloud.mcp:mcp-toolbox-sdk-java:1.0.0). Credentials now resolve on every request, so tokens can rotate, and server-bound parameters such as tenant_id are stripped from tool definitions so the model cannot set them. The SDK also warns when credentials would travel over plain HTTP. (official)

Letting a model fill in a tenant ID is a classic multi-tenant hole: one persuasive prompt and it queries someone else’s rows.

For Software Developers: Move tenant_id and similar values out of the tool parameters the model sees and bind them server-side with the 1.0 SDK, so there is no field left for the model to set.

AI-assisted SDLC

[2026-10-07] A Coleman Parkes survey of 300 senior engineering leaders, run for debugging vendor Undo, found teams spend 16.9 hours a week debugging against 9.8 writing code. 35% of generated code reaches production before the team fully understands it, and 79% say agents write code faster while releases ship no sooner. (source)

Undo sells debugging tools, so the hours figure comes from a party with a stake in it. The 35% figure on code shipped before it is understood is the one worth raising at a team retro.
[2026-10-06] JetBrains Research and Lund University, drawing on workshops with 17 practitioners and a survey of 43, argue that reviewing AI-generated code fails on effort allocation: the diff looks equally confident everywhere. They propose overview, then file-level risk ranking, then snippet detail, and warn that tools which only aid comprehension speed up low-risk reviews and nothing else. (official)

Teams can try the ordering without new tooling: skim the whole change first, rank files by what they touch, and spend the reading time on the risky ones.
[2026-10-06] GitHub says pushes grew 4.9 times in a year to 3.35 billion a month, with agents checkpointing after nearly every action. It is moving repository data to Azure Blob Storage behind cache-like compute and reports up to 35 times the write throughput in internal benchmarks. (official)

Anyone running a self-hosted Git server for agent-heavy teams should expect the same push curve, without GitHub’s storage rebuild behind it.

AI cost tracking & telemetry

[2026-10-07] Anthropic is adding monthly API credits to Claude subscriptions: $100 on Max 5x, $200 on Max 20x, and $20 or $100 per Team seat, pooled up to $500. They cover the API, Managed Agents and the Agent SDK, not interactive Claude Code or cloud marketplaces, and expire each billing cycle. Free, Pro and Enterprise get none. (official)

Because the credits lapse each cycle, they reward steady API use over occasional bursts; Pro subscribers who build on the API get nothing from this.

Practice & craft

[2026-10-07] Stack Overflow’s guide to evals as a deployment gate fails CI below a 98% pass rate, gates each slice separately because 99% overall can hide a 70% slice, and flags drift only when a metric worsens past a set minimum with |z| above 2. About 5% of live decisions are shadow-scored to feed new golden cases. (official)

Per-slice gates are the piece most eval setups leave out, and without them an overall pass rate can stay green while one segment fails.
[2026-10-07] Lumen’s Black Lotus Labs tracked PoeLLM, a cryptomining campaign that has hit more than 3,400 servers since April, including exposed LiteLLM gateways, then turns victims into scanners. Its command server’s address is hidden in a poem in a GitHub repository. (source)

An exposed LiteLLM gateway usually holds provider API keys, which makes it worth more to an attacker than the CPU it mines with.

Research worth reading

[2026-10-07] PackHallu hides prompts in ordinary rule files such as AGENTS.md or .cursorrules that steer coding agents into swapping real dependencies for attacker-controlled packages, evolving the prompts from agent trajectories. The authors report high success that transfers across models and agent frameworks; the abstract gives no figures. Rule files from outside contributors deserve the same review as code. (paper)

Dependency changes in an agent-written diff are worth checking against the previous lockfile before merge, whoever wrote the rule files.
[2026-10-07] A study of LLM vulnerability-patching benchmarks, built on 112 historical bugs from 84 C/C++, Go and Rust projects, finds patches that pass the proof-of-concept test often fail the developers’ own tests. Pass rates on those barely move with newer models, and harness details such as timeouts and permissions shift scores. (paper)

An agent’s security fix needs the project’s full test suite, not just the exploit reproduction, before anyone calls it done.

Watch list

Mistral Large 4 weights: preview only; weights promised by the end of October, license unnamed.
Step 5 Preview’s weights: not uploaded; due October 15.
Decisions API general availability: public beta since October 6; OpenAI says “the coming weeks”.
Gemini 4 Argon for paid API users: Fairwind cohort only; no date. Reflection Beam’s weights, still unreleased, drop off until there’s news.