Claude Haiku 5.5 lands at $0.10 input (90% cheaper), Claude for Google Workspace opens in beta, Anthropic Cyber Mission launches with free OSS Scanner

Claude Haiku 5.5 shipped October 7 as Anthropic's fastest, cheapest, and most capable small model to date. For prompts up to 100K tokens: $0.10 input, $0.50 output, $0.01 cache reads — 90% cheaper than Haiku 4.5's $1/$5/$0.10. Above 100K tokens: $0.50 input, $2.50 output, $0.05 cache reads — 50% cheaper. Cache writes: $0.125 (≤100K) and $0.625 (>100K) per MTok. Anthropic estimates average cost around 75% lower than Haiku 4.5. Context window is 1M tokens. It's the first Haiku-class model with an adjustable effort setting, letting teams trade cost against intelligence. Benchmark leaps over Haiku 4.5 are dramatic: GDPval-AA 1620 vs 735, OSWorld 2.1 72.4% vs 15.7%, Humanity's Last Exam with tools 57.4% vs 18.7%, Terminal-Bench 39.2% vs 0%. Haiku 5.5 also outperforms GPT-6 Luna on GDPval-AA (1620 vs 1437), OSWorld (72.4% vs 48.9%), and Terminal-Bench (39.2% vs 16.4%). Customer results: Asana saw 30%+ lower latency and up to 2.5x faster inference per agent turn; HubSpot scored 92.8% on its CRM suite; AlphaSense 0.84 vs 0.76 across 400 queries. Best for: summaries, compaction, database queries, classification, subagent work, live customer support, and browser use. Model ID: claude-haiku-5-5. Available on AWS, Google Cloud, Azure, and Claude Platform.

Alongside Haiku 5.5, Anthropic halved Sonnet 5.5's cache-read price from $0.20 to $0.10 per million tokens. Anthropic estimates the reduction cuts most agentic Sonnet 5.5 workloads by around 20% without any code changes. No change to base input ($2) or output ($10) pricing. The Haiku 5.5 launch also brought monthly API credits for Max and Team plan subscribers: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team.

Claude for Google Workspace opened as a public beta on October 6, available on every paid plan (Pro, Max, Team, Enterprise). One install from the Google Workspace Marketplace adds a Claude sidebar to Google Docs, Sheets, and Slides simultaneously. In Docs, the sidebar reads the open file and current selection, fixes or restyles text while preserving formatting, and proposes larger rewrites as suggestion cards. In Sheets, it writes formulas, builds pivot tables, native charts, and new tabs; for joins or cleaning it pulls ranges into Python and writes results back. In Slides, it builds slides using the deck's own layouts and themes and flags overlapping or off-slide elements. Default mode is 'Ask before edits'; 'Accept all edits' allows unattended runs. Beta connectors also launched so Claude can create and live-edit Google files from chat — access follows existing Google sharing permissions. Enterprise controls include Compliance API, CMEK, and OpenTelemetry audit export. Beta limitations: six-minute max per operation (Google limit), no Firefox support, closing the sidebar stops a running task, no third-party inference (Bedrock, Vertex, Azure AI Foundry).

Two new beta artifact types launched October 8. Claude Dashboards (paid plans): ask plain-language questions and Claude builds live-updating dashboards connected to BigQuery, Redshift, ClickHouse, Databricks, Snowflake, and Salesforce — each chart shows its underlying query and last-refresh time. Export to Amplitude, Grafana, Hex, Mixpanel, Omni, Perplexity, PostHog, or Sigma; Looker, monday.com, and Tableau coming. Claude Motion (Team and Enterprise): type /motion and Claude writes editable code that animates text, charts, shapes, and images as a short MP4 — no AI-generated footage or people. Export to Adobe, Descript, HeyGen, Higgsfield, invideo, Luma AI, and Runway. Both are off by default — enable in Organization settings under Artifacts. Separately, Claude Docs, Slides, and Design graduated from beta to GA on all plans including Free, adding CMEK and admin-controlled artifact templates. Docs, Slides, and Design turn on automatically for all organizations on October 15 and can be enabled earlier in settings.

Anthropic launched the Cyber Mission on October 8, framed as a long-term commitment to securing critical infrastructure and open-source software. The Critical Infrastructure Defense Program works through security vendors and OT integrators — not directly with operators — giving a small initial cohort frontier Claude models, on-site Anthropic engineers, and threat research. Eleven founding partners: Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC, and Rockwell Automation. Other providers can register interest. The OSS Scanner is a free, opt-in service for established open-source projects: Anthropic's strongest models periodically scan enrolled repos and email emailed reports with proof of concept, explanation, and suggested fix. Reports are sent without human review — Anthropic expects a 90%+ true-positive rate but notes some errors are possible and there is no 90-day disclosure deadline for unvalidated findings. Enroll by opening a PR to anthropics/oss-scanner adding a project.yaml with repo, contact, and Dockerfile. Costs are covered through Anthropic's Defender Advantage Fund. Booz Allen and Comcast also published customer stories this week on using Claude Mythos models through Project Glasswing for vulnerability discovery and validation.

The Cyber Verification Program expanded on October 6 with three access tiers for qualifying security professionals: Defense (for defenders working on critical systems), Red Team (for offensive security researchers), and Specialized (for teams requiring both Mythos-level access and custom safeguard configurations). Separately, the Models API now returns a capabilities.server_tools field on each model indicating whether it supports web_search and code_execution — useful for feature-flagging model-specific behavior in multi-model pipelines without hardcoding model names. Dynamic Workflows for Claude Managed Agents also entered beta on October 9 (requires managed-agents-2026-04-01 beta header): agents can now write and run multi-phase background workflows — such as reviewing hundreds of documents — without pausing for a human to advance each phase.

Claude Code shipped four versions this week. v2.1.292 (Oct 6): plugin install from the marketplace in Claude Code; effort parameter for subagents to tune thinking depth per spawned agent; fixes across subagents, plugins, Chrome, and Slack. v2.1.293 (Oct 7): Haiku 5.5 becomes the default Haiku model; reliability fixes. v2.1.296 (Oct 10): controls for agents, workflows, and gateways; larger Read tool limit for taking in bigger files; subagent auto-compact to manage context in long-running background agents; reliability and bug fixes across cloud sessions, VS Code, Slack, Windows, and self-hosted runners.