Claude Latest News
Curated updates on Claude products, API changes, regional policy, and community-reported security trends.
Items are manually curated from official Anthropic announcements and community reports. Links open external sources.
Claude Latest News
Curated updates on Claude products, API changes, regional policy, and community-reported security trends.
Claude Code 2.1.292–2.1.294: permission and sandbox fixes, stdio MCP protocol negotiation, prompt-hook fix
Claude Code 2.1.292 fixes several permission bypasses: PreToolUse hook approvals and auto mode skipped the permission prompt for file reads from network (UNC) paths, sandboxed commands could read the staged copies of `/ultrareview` uploads, and a tampered on-disk cache of server-managed settings could switch off the built-in policy plugin while the settings fetch failed. It also makes local (stdio) MCP servers negotiate protocol version 2026-07-28 by default on every install, including Bedrock, Vertex, and Foundry; set `MCP_PROTOCOL_NEGOTIATION=legacy` to opt out. Version 2.1.293 sets Claude Haiku 5.5 as the default Haiku model. Version 2.1.294 fixes `prompt` and `agent` hooks written as instructions (for example "Block commands that…") so they block what they should, and improves how Stop and SubagentStop prompt hooks are judged.
Takeaway:If you rely on `prompt` or `agent` hooks to block commands, upgrade to 2.1.294 and re-test them; earlier versions could let blocked commands through when the hook was phrased as an instruction.Source: Anthropic ↗Related Guide →Claude Haiku 5.5 launches at $0.10/$0.50; Sonnet 5.5 cache reads cut to $0.10; Max and Team get monthly API credits
Anthropic released Claude Haiku 5.5 (`claude-haiku-5-5`): 1M context, 128k max output, adaptive thinking with an `effort` parameter, and $0.10 input / $0.50 output per MTok for prompts up to 100k tokens ($0.50 / $2.50 above 100k). Anthropic says it costs about 75% less than Haiku 4.5 on average and targets subagents, summarization, compaction, and browser use; Sonnet 5.5 and Opus 5.5 remain the recommended models for complex agentic coding. Code written for Haiku 4.5 can break: manual extended thinking (`budget_tokens`) returns a 400 error, adaptive thinking is on by default, and a new tokenizer counts more tokens for the same text. The same day, Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per MTok, and Claude Max (5x: $100, 20x: $200 per month) and Team (up to $500 per month, pooled) began receiving monthly Claude Platform API credits. The Python and TypeScript SDKs also add beta classes for the browser use and computer use tools.
Takeaway:Move cheap subagent and summarization calls to Haiku 5.5 only after removing `budget_tokens` and re-checking token counts under the new tokenizer. Max and Team admins can claim the API credits following the Help Center article.Source: Anthropic ↗Related Guide →Managed Agents `web_fetch` now fetches only URLs already seen in the session
To reduce data-exfiltration risk, `web_fetch` in Claude Managed Agents fetches only URLs that already appeared in the session: in a user message, in a `web_search` result, or in a page fetched earlier. A URL that appears only in Claude's own output, the system prompt, an attached document, or tool output such as `bash`, `read`, or an MCP tool returns a `url_not_in_prior_context` error. Separately, in cloud environments with `limited` networking, `allowed_hosts` now also applies to `web_search` and `web_fetch`. Creating or updating a session fails with a 400 error when an enabled web tool's `allowed_domains` has an entry outside `allowed_hosts`. An `allowed_hosts` entry matches one exact host unless it starts with `*.`.
Takeaway:Send any URL the agent must fetch in a `user.message` event, and list each host in `allowed_hosts` — `docs.example.com` is not covered by `example.com`.Source: Anthropic Docs ↗Related Guide →Cyber Verification Program expands to three access tiers covering Opus 5.5, Sonnet 5.5, and Mythos 5.1
Anthropic merged Project Glasswing and the Cyber Verification Program (CVP) into one program with three tiers. Defense Access covers SOC and incident-response work, malware reverse engineering, and vulnerability analysis; Anthropic aims to respond within a few days. Red Team Access adds authorized penetration testing and red-teaming for organizations only, with a review of a few weeks; applicants are enrolled in Defense Access in the meantime. Specialized Access has the fewest blocks and is reserved for verified organizations authorized to test systems such as power grids and interbank transfer infrastructure, reviewed together with the US government. Enrolled organizations must allow data retention; Enterprise Frontier Safeguards, planned for later this fall, will allow storage in customer-controlled cloud. Generally available models keep conservative cyber safeguards and still handle code review, patching known issues, and alert triage. Existing CVP members keep their settings and are evaluated automatically for the new models.
Takeaway:Defensive security work blocked on generally available models is an eligibility question, not a prompt-wording problem: apply for the matching tier instead of rewording requests.Source: Anthropic ↗Related Guide →Hong Kong users report Claude account suspensions from Oct 1; Anthropic has not commented
From October 1, 2026, Hong Kong users reported "Account Suspended" or "Access Denied" screens on claude.ai, including paid Pro subscribers who had connected through VPN exits in supported regions such as Singapore. Hong Kong, mainland China, and Macau are not on Anthropic's Supported Regions list; Taiwan, Singapore, and Japan are. The Help Center lists "account creation from an unsupported location" as a ban reason, and a VPN exit in a supported country does not change eligibility. One posted notice cited suspicious signals and a Usage Policy violation without naming a trigger. Media and community analysis points to shared data-center VPN IPs, virtual-card or unsupported-region payments, and frequent country switching, but Anthropic has not confirmed any of this and the number of affected accounts is unknown. One user reported that a data export returned account information without conversation history; the Help Center says exportable data can be restricted depending on the violation.
Takeaway:A supported-country VPN exit does not make an unsupported-region account eligible. If suspended: export data first, appeal with factual details at claude.ai/restricted, and do not register replacement accounts while the appeal is pending.Claude Sonnet 4.5 deprecated with Nov 30 retirement; Admin API methods leave beta in SDKs and the `ant` CLI
Anthropic announced the deprecation of Claude Sonnet 4.5 (`claude-sonnet-4-5-20250929`), with retirement on the Claude API scheduled for November 30, 2026, and recommends migrating to Claude Sonnet 5.5. In Python SDK 1.10.0, TypeScript SDK 0.130.0, C# SDK 12.52.0, Go SDK 1.77.0, Java SDK 2.67.0, PHP SDK 0.53.0, Ruby SDK 1.75.0, and `ant` CLI 1.37.0, the Admin API methods for organization info, members, invites, workspaces, API keys, rate limits, service accounts, workload identity federation, customer-managed encryption keys, and compliance settings are out of beta under `client.organization` and `ant organization`. The beta paths remain available.
Takeaway:Replace hard-coded `claude-sonnet-4-5-20250929` before November 30, and read the Sonnet 5.5 migration guide first: `thinking` and forced `tool_choice` behave differently from earlier models.Source: Anthropic Docs ↗Related Guide →Claude Sonnet 5.5 launches at an unchanged $2/$10, with thinking blocks bound to the account
Anthropic released Claude Sonnet 5.5 (`claude-sonnet-5-5`), the second Claude 5.5 model after Opus 5.5: 1M context, 128k max output, 30%+ faster than Sonnet 5, and up to 30% cheaper on most work because it uses fewer tokens at the same $2 / $10 per MTok. Migrating from Sonnet 5 breaks in five places: up-front thinking is turned off with `thinking: {"type": "between_tools"}` instead of `"disabled"` (at `high` effort or below); `tool_choice` types `any` and `tool` return 400; thinking blocks are tied to the model and conversation; the earlier `computer_20251124` tool is rejected on the Claude API and Google Cloud; and the advisor tool rejects Opus 4.8, Opus 4.7, and Sonnet 5 as advisors. Thinking blocks that Sonnet 5.5 produces work only in the account that produced them or an account linked to it. When another account sends one, the API drops the block before the model sees it and the request still succeeds. Claude Code 2.1.284 makes `sonnet` resolve to Sonnet 5.5 on the Anthropic API.
Takeaway:An API relay or gateway that rotates several accounts within one conversation can silently lose Sonnet 5.5 thinking blocks. Keep each conversation on a single account.Source: Anthropic ↗Related Guide →Claude Code 2.1.283 adds enterprise model allow/deny lists and third-party auto mode default
Claude Code 2.1.283 introduces managed `availableModelsMatch` (`exact` pins a listed model version until admins update the list) and `deniedModels` (blocks specific models even when otherwise allowed). Third-party API, Vertex, Bedrock, or Foundry sessions with telemetry off now start in auto mode when no `permissions.defaultMode` is set. Other highlights: `/doctor prompt-audit` scans CLAUDE.md and skills for stale patterns, MCP tool images are saved to disk for follow-up tools, and `claude plugin validate` rejects install-unsafe marketplace names.
Takeaway:Enterprise admins should pair `deniedModels` with explicit `permissions.defaultMode` on sensitive repos — auto mode now applies outside claude.ai subscriptions too.Source: Anthropic ↗Related Guide →Anthropic resumes billing for pre-output refusals in bio, frontier LLM, and reasoning-extraction categories
From September 24, 2026, Messages API requests that refuse before any output are billed again when `stop_details.category` is `bio`, `frontier_llm`, or `reasoning_extraction` — the categories Anthropic says have the lowest false-positive rates. Mid-stream refusals were already charged. Other pre-output refusal categories stay free; fallback credits are unchanged. Charges use the executing model’s normal rates on all platforms.
Takeaway:Budget for classifier-triggered hard stops on biology- or capability-probing prompts — they can cost full input tokens even with empty assistant output.Source: Anthropic Docs ↗Related Guide →Compliance API Activity Feed strips filenames and artifact titles from events
The Compliance API Activity Feed no longer exposes file names, project document names, or artifact titles. The `filename` and `title` fields on file, project-document, and artifact activities are always empty or omitted, including for historical records. Enterprise teams that need human-readable names must resolve IDs with a Compliance Access Key scoped to `read:compliance_user_data`.
Takeaway:Update SIEM parsers and audit dashboards that relied on inline titles — store only activity IDs unless you call the compliance user-data endpoints.Source: Anthropic Docs ↗Related Guide →Prompt cache diagnostics exit beta on the Messages API
Cache diagnostics is GA on the Claude API and no longer needs the `cache-diagnosis-2026-04-07` beta header. Opt in by sending a `diagnostics` object on Messages requests; every `POST /v1/messages` response now includes a `diagnostics` field (`null` when you did not opt in). Requests that still send the old beta header behave as before.
Takeaway:Send `diagnostics` on every turn in multi-step agents — header-only requests no longer store fingerprints for later `previous_message_id` checks.Source: Anthropic Docs ↗Related Guide →Claude Opus 5.5 launches at $4/$20 with ~40% lower typical cost than Opus 5
Anthropic released Claude Opus 5.5 (`claude-opus-5-5`) as the first Claude 5.5-family model: 1M context, 128k max output, always-on adaptive thinking, priced at $4 / $20 per MTok with cache reads at $0.20 / MTok (vs $5 / $25 and $0.50 on Opus 5). Anthropic positions it near Claude Fable 5.1 on most agentic coding and knowledge work while using fewer tokens per task. API constraints mirror Fable 5.1: `thinking` cannot be disabled, and `tool_choice` types `any`/`tool` return 400. Fast mode is available in research preview. Pro, Max, Team, and seat-based Enterprise plans get higher five-hour limits plus a user-chosen rate-limit reset.
Takeaway:Migrate long-horizon coding agents from Opus 5 for cost; drop disabled-thinking and forced-tool `tool_choice` patterns before switching model IDs.Source: Anthropic ↗Related Guide →Claude Code 2.1.277 reads AGENTS.md, removes TaskOutput, and hardens headless sessions
Claude Code 2.1.277 uses `AGENTS.md` as project instructions when no `CLAUDE.md` exists (configurable under “Project instructions”; not yet on Bedrock, Vertex, or Foundry). The deprecated TaskOutput tool is removed — use Read on background task output files instead. Gateway egress can set `CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1` so proxies receive hostnames instead of local DNS resolution. Dozens of fixes cover `claude -p` hangs, empty text-block resume failures, plugin reinstall races, and headless cost totals resetting to zero on resume.
Takeaway:Standardize on either CLAUDE.md or AGENTS.md per repo — mixed or empty instruction files now change which rules Claude loads at session start.Source: Anthropic ↗Related Guide →Life Sciences Verification Program opens for vetted biology workloads on Mythos, Opus, and Sonnet
Anthropic opened the Life Sciences Verification Program (LSVP) for teams vetted on credentials, security posture, and research oversight. Standard Use grants cover most R&D workflows on Mythos 5.1, Opus 5, Sonnet 5, and future models via API, Claude Science, claude.ai, and Claude Code, with classifiers tuned for legitimate biology. High-risk Use add-ons remove life-science blocks for single projects on six-month renewal. Enforcement shifts toward offline monitoring with mandatory 30-day retention on LSVP traffic; cyber classifiers stay active.
Takeaway:Routine biology Q&A blocked on consumer Fable is not a bypass bug — legitimate labs need LSVP verification instead of prompt jailbreaks.Source: Anthropic ↗On-demand conversation compaction enters beta on the Messages API
With the `compact-2026-09-04` beta header, Messages API callers can compact a thread on demand via the top-level `compaction` parameter. The API returns a signed `compaction` block summarizing the messages you sent; later turns send that block instead of the full history. You choose when to compact, can run compaction in the background, and may keep recent turns verbatim after the summary — preserved thinking in those turns can remain valid on supported models.
Takeaway:Use compaction for long agent loops instead of hand-rolling summaries — unsigned DIY summaries still break thinking-block binding on post–Aug 31 accounts.Source: Anthropic Docs ↗Related Guide →September 2026 threat report: agentic cyber ops, distillation, and Claude Code in state campaigns
Anthropic’s September 2026 Threat Intelligence report covers misuse disrupted from December 2025 through August 2026 across cyber operations, influence ops, scams, surveillance, biology, weapons, and distillation. Haiku, Sonnet, and Opus were abused; Fable and Mythos-class models were not, except one illicit distillation case. Case studies describe multi-agent kill chains — including operators refining Claude Code skills to rebuild malware when detections fire — and reinforce that high-volume automated extraction remains an enforcement priority.
Takeaway:Treat Claude Code auto mode and stolen API keys as high-risk surfaces — Anthropic explicitly links skill-driven workflows to espionage-scale operations.Claude Managed Agents gain server-side `auto` permission evaluation; `ant beta:sessions connect` ships
Managed Agents permission policies now accept `auto`: the server evaluates each agent or MCP tool call and runs, denies, or pauses for approval. `agent.tool_use` and `agent.mcp_tool_use` events include an `evaluation` field beside `evaluated_permission`. Anthropic CLI 1.32.0 adds `ant beta:sessions connect` to attach a terminal to a Managed Agents session, send messages, and approve pending tool calls; pass `--web` to open the Console session viewer locally.
Takeaway:Prefer server-side `auto` over client-only allowlists when running untrusted MCP tools in managed sandboxes.Source: Anthropic Docs ↗Related Guide →Official disabled-account appeal portal live at claude.ai/restricted
Anthropic now routes disabled-account and organization-ban appeals through a dedicated form at https://claude.ai/restricted instead of generic support tickets. The form collects account email, ban context, and usage justification. Community reports in September 2026 show successful reinstatements within 1–3 business days when appeals cite legitimate software engineering or academic use with verifiable details.
Takeaway:Use claude.ai/restricted for Disabled status; reserve usersafety@anthropic.com for API-tier warnings and false positives.Source: Anthropic ↗Related Guide →Claude Code 2.1.259 adds org-managed MCP servers and headless permission controls
Claude Code 2.1.259 introduces `managedMcpServers`: organizations can provision HTTP/SSE MCP servers to every user (same shape as `.mcp.json`; command-based entries are skipped). `allowedMcpServers` now governs only user-added servers. For CI and unattended hosts, `--permission-prompts none` auto-denies anything that would prompt while the active permission mode (including auto mode) still applies. GitLab merge-request commands (`glab mr create/merge/close/reopen/note/update`) are recognized in tool summaries.
Takeaway:Enterprise admins should centralize approved MCP endpoints via managed settings instead of relying on per-user `.mcp.json` copies.Source: Anthropic ↗Related Guide →Accounts created after Aug 31, 2026 face stricter thinking-block replay binding
Since August 31, 2026, accounts created on or after that date are subject to tighter binding on adaptive thinking blocks across multi-turn API sessions. Replaying, reordering, or tampering with thinking blocks in stored context is treated as a distillation signal and can trigger organization-level disable. This aligns with Anthropic's ongoing crackdown on model distillation and automated capability extraction.
Takeaway:Never cache and replay raw thinking blocks in agent pipelines; store only final assistant text and re-derive reasoning on each turn.Source: Anthropic Docs ↗Related Guide →Claude Fable 5.1 and Mythos 5.1 launch with lower cache-read pricing
Anthropic released Claude Fable 5.1 (`claude-fable-5-1`) and Claude Mythos 5.1 (`claude-mythos-5-1`): 1M context, 128k max output, always-on adaptive thinking, at $10/$50 per MTok with cache reads cut to $0.25/MTok (0.025× input vs 0.1× on other models). Fable 5.1 is GA on the API and cloud partners; Mythos 5.1 stays on trusted-access programs. API changes include no `tool_choice` types `any`/`tool`, stricter thinking-block replay binding for accounts created on or after Aug 31, 2026, and beta per-message effort and turn-scoped system messages. Claude Code 2.1.257 sets Fable 5.1 as the default Fable model.
Takeaway:Agent workloads with heavy cache hits benefit most from Fable 5.1; audit `tool_choice` and multi-turn thinking replay before migrating production pipelines.Source: Anthropic Docs ↗Related Guide →Model Hardware Standard research preview opens for labs and manufacturers
Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification for AI agents to operate physical devices such as microscopes, liquid handlers, and robotic arms. MHS is model-agnostic and reachable via MCP, CLI, or code APIs; access is invite-only for scientific research labs and advanced manufacturers ahead of an open-source release. Early partners report faster device integration, quicker experiment iteration, and live fault detection.
Takeaway:MHS is not yet open source; lab and manufacturing teams should apply at modelhardwarestandard.com rather than building ad hoc device drivers.Source: Anthropic ↗Files & Skills SDK paths go GA; personal and service account API keys launch
In Python SDK 1.2.0, TypeScript SDK 0.122.0, and matching Go/Java/Ruby/C# releases, `client.beta.files` and `client.beta.skills` no longer send beta headers and mirror `client.files` and `client.skills`. Beta `BetaSkill` is renamed `BetaContainerSkill`, and `client.beta.skills.delete()` now removes a Skill and all its versions. The Claude Console also supports personal keys and service account keys scoped to workspaces or admin endpoints.
Takeaway:Remove `files-api-2025-04-14` and `skills-2025-10-02` beta headers from production SDK calls and migrate to the stable client paths.Source: Anthropic Docs ↗Related Guide →Cowork built-in browser ships; Claude in Chrome reaches GA with autonomous actions
Claude Cowork on desktop now opens a dedicated side-panel browser for web tasks—no extension required; it rolls out to Pro, Max, and Team on macOS, Windows, and Linux (beta), with Enterprise admins controlling access in Organization settings. Separately, Claude in Chrome is GA on all paid plans: Claude can read pages, navigate, and act autonomously in the user's Chrome session, with a safety classifier validating each action. Enterprise Chrome extension defaults flip to enabled on Sep 10, 2026 unless already disabled.
Takeaway:Pick built-in browser for isolated web handoffs and Claude in Chrome when you need existing logged-in sessions; review org browser policies before Sep 10.Source: Anthropic ↗Related Guide →Compliance API session endpoints exit beta; Admin API arrives in SDKs and ant CLI
Cowork and Claude Code Compliance API session endpoints are now GA. Local session endpoints also return Claude Science and Claude for Microsoft 365 transcripts (beta, Enterprise). The Admin API is available in the `ant` CLI and Python, TypeScript, C#, Go, Java, PHP, and Ruby SDKs under `client.beta.organization`, covering members, workspaces, API keys, rate limits, and service accounts.
Takeaway:Enterprise admins can automate org governance via SDK Admin API calls instead of curl-only workflows.Source: Anthropic Docs ↗Related Guide →Claude Cowork memory syncs across chat; topics editable with sensitive-topic controls
Memory now works across chat and Cowork in the cloud. All remembered items appear under Topics in Settings > Memory, where users can edit or delete entries. Health- or belief-related topics stay out of memory unless "Include sensitive topics" is enabled. Memory is on by default for Free, Pro, and Max; off by default for Team and Enterprise.
Takeaway:Review your organization memory defaults before rolling Cowork to Team or Enterprise users.Source: Anthropic Help Center ↗Python SDK v1.0 ships with httpx2 migration and legacy API removal
Anthropic released Python SDK v1.0: the HTTP layer moves from httpx to the maintained httpx2 fork (requires Python 3.10+). Long-deprecated surface is removed, including the legacy Text Completions API, `temperature`/`top_p`/`top_k` on Messages, and the tool runner's client-side `compaction_control`. Async `.with_raw_response` now needs `await response.parse()`. See the v1 migration guide for full breaking changes.
Takeaway:Upgrade older pipelines from /v1/complete to /v1/messages before updating the SDK dependency in production.Source: Anthropic Docs ↗Related Guide →Computer use, browser use, Files API, and Agent Skills exit beta
Anthropic graduated four major platform features: computer use (`computer_toolset_20260801`) and browser use (`browser_toolset_20260801`) no longer need beta headers; Files API and Agent Skills (`/v1/skills`) are GA too. All four are available on Fable 5, Mythos 5, Opus 5, Sonnet 5, and Opus 4.8.
Takeaway:Beta request headers can now be removed from production workloads invoking computer use and tool capabilities.Source: Anthropic Docs ↗Related Guide →Claude text watermarking rolls out globally for EU AI Act compliance
New Claude models embed an invisible SynthID-Text statistical watermark in generated text, applied globally (not EU-only) to meet AI transparency regulations. The mark carries no user or org identifiers and adds no extra tokens; detection and verification tools are now standardized.
Takeaway:Review our watermarking detection guide for techniques on identifying and safely handling generated content.Source: Anthropic ↗Related Guide →Claude Code auto mode becomes default on Pro, Max, and Team plans
Since Aug 14, new Claude Code sessions on Pro, Max, and Team run in auto mode — a background classifier approves routine tool calls instead of prompting every command. Classifier token overhead is no longer charged on these plans; Enterprise and API access remain opt-in.
Takeaway:Pin `permissions.defaultMode` in `~/.claude/settings.json` if you need manual approval on sensitive repositories.Source: Anthropic ↗Related Guide →Claude Sonnet standard pricing ($2/$10) made permanent
Anthropic cancelled the planned price increase to $3/$15 per million tokens. Sonnet stays at $2 input / $10 output per MTok as the permanent standard API price, offering unbeatable cost-performance for developer workflows.
Takeaway:Combine Sonnet with prompt caching to reduce effective input costs down to $0.20 per million tokens on cache hits.Source: Anthropic Docs ↗Related Guide →Fable 5 biology safeguards tuned; fallback false positives down ~85%
Anthropic retuned Fable 5's biology safety classifiers so benign health and education queries trigger fewer fallbacks to Opus 5 — about 85% fewer biology-related fallbacks in testing. Dual-use requests (virology, toxicology, molecular design) still route to the less capable model; trusted-access pathways for frontier biology remain in development.
Takeaway:Educational biology questions no longer suffer artificial performance throttling or unexpected model switching.Source: Anthropic ↗Related Guide →Enterprise skill and plugin security scanning enters beta
Enterprise plans can enable skill and plugin security scanning to automatically check third-party skills and plugins for malicious content when someone uploads or edits them.
Takeaway:Enable scanning in org settings before allowing members to install community skills in production workspaces.Claude Opus 4.1 retired; migrate to Opus 4.8 or Opus 5
Anthropic ended support for claude-opus-4-1-20250805 after a 60-day deprecation window. API calls to the old model ID now return errors; Opus 4.8 and Opus 5 are the recommended replacements.
Takeaway:Check your API configuration files and update hardcoded model strings to prevent runtime 400 errors.Source: Anthropic Docs ↗Related Guide →Anthropic releases Claude Opus 5 as default on Pro and Max
Claude Opus 5 (`claude-opus-5`) delivers near-Fable 5 intelligence at half the price, with 1M context, 128k max output, and adaptive thinking on by default. It is the default on Claude Max and the strongest model on Pro, priced at $5/$25 per MTok. Beta features include mid-conversation tool changes and automatic API fallbacks for safety-classifier refusals.
Takeaway:Enable `fallbacks` with `"default"` mode (beta) so flagged Opus 5 requests route to another model instead of hard-blocking.Source: Anthropic ↗Related Guide →Anthropic H1 2026 Transparency Report: 11.4M accounts banned; ~10.5% appeal success rate
Anthropic's Transparency Hub H1 2026 report shows 11.4 million policy-violating accounts disabled in six months, 398,000 appeals received, and 42,000 successful reinstatements (~10.5% success rate). Primary enforcement targets: unsupported-region access, abuse detection, multi-account bulk registration, and model distillation attacks. Disabled accounts should use the dedicated appeal form at claude.ai/restricted; API false-positive or violation warnings go to usersafety@anthropic.com.
Takeaway:Appeals with clear legitimate-use proof have a real chance of reinstatement — do not panic-create replacement accounts while an appeal is pending.Anthropic launches Claude Sonnet 5 with adaptive thinking and 1M context
Claude Sonnet 5 (`claude-sonnet-5`) is a drop-in upgrade over Sonnet 4.6 with 1M-token context, 128k max output, and adaptive thinking on by default. Manual extended thinking and non-default `temperature`/`top_p`/`top_k` now return 400 errors. A new tokenizer produces ~30% more tokens for the same text. Launch pricing was $2/$10 per MTok, later made permanent on Aug 10.
Takeaway:Remove `thinking: {type: "enabled"}` and sampling params when migrating from Sonnet 4.6; use the `effort` parameter instead.Source: Anthropic ↗Related Guide →Anthropic launches Claude Fable 5 and restricted Claude Mythos 5
Claude Fable 5 (`claude-fable-5`) is Anthropic's most capable widely released model with 1M context, always-on adaptive thinking, and safety classifiers that can return `stop_reason: "refusal"`. Claude Mythos 5 shares the same base model with lifted safeguards for Project Glasswing partners. Access was briefly suspended June 12–July 1 due to export controls before global redeployment.
Takeaway:Use the opt-in `fallbacks` parameter (beta) to route refused Fable 5 requests to another model instead of hard-blocking.Source: Anthropic ↗Related Guide →
Community Anti-Ban Field Reports
Real-world tips from X (Twitter) on Claude bans, appeals, payments, and Claude Code — curated with direct links to original posts.
Long-form anti-ban guide after the National Day ban reports: environment, payment, usage, local settings
An X article organized in four layers. Environment: one fixed exit node per account, TUN mode, system timezone matched to the exit region, and WebRTC, DNS, and leftover-proxy checks. Account and payment: App Store in-app purchase as a refund fallback, and ramping from Free to Pro before Max. Usage: do not drain the weekly quota. Claude Code: official client only, with telemetry switches set in `settings.json`. It ends with an 11-step minimum setup and an ordered checklist for appeals and refunds. The post contains affiliate links for third-party IP and card services; this site does not endorse them.
Takeaway:Fix one exit node and one device per account, and check WebRTC, DNS, and leftover proxy variables (`ANTHROPIC_BASE_URL`, `HTTPS_PROXY`) before every login. Keep code in local Git so a ban does not take project state with it.Related guide: Claude Anti-Ban Essentials: Quick Safety Checklist →
View original on X (Twitter) ↗Opens external link in a new tabTwo accounts banned on the same day with different notices: network is not the only trigger
After sharing anti-ban tips with a friend, the author lost two accounts on the same day. One notice said the login country was wrong; the other cited a Usage Policy violation. The author says they regularly use Claude for borderline tasks such as decompilation and adult-content research, and suspects content, not only network, contributed to the ban.
Takeaway:Bans arrive through two separate channels: access signals (region, login country) and content (Usage Policy). A clean network setup does not protect an account whose prompts trigger policy classifiers.Related guide: Claude Ban Case Studies & Root Cause Analysis →
View original on X (Twitter) ↗Opens external link in a new tabSix-month stable setup, shared during the National Day ban reports
Shares a configuration used for six months without a ban: a dedicated Mac, US-West timezone, one account that was never switched, Surge in rule mode, a dedicated residential IP in Hawaii, a US credit card, 50–80% weekly quota use, and remote access to the host when away. The author adds that no method is 100% safe and that a stable IP plus moderate usage most likely lowers the risk.
Takeaway:Keep it boring: one account, one device, one exit IP, and a weekly quota that stays below the limit.Related guide: OS & Browser Environment Cleanup & Residential IP Guide →
View original on X (Twitter) ↗Opens external link in a new tabThree-layer anti-ban review from dozens of field reports: leaks, billing credit, usage pacing
Summarizes feedback from dozens of users. Many bans trace to DNS or WebRTC exposing the real location, not to the IP type. Billing credit weighs as much as network: web-page billing with domestic dual-currency or virtual cards is a frequent ban pattern, while an App Store in-app purchase with a US Apple ID hides issuer details. New accounts need a ramp-up period, and the weekly quota should not be exhausted every cycle. The author says language is not the deciding factor; network fingerprint and call frequency are.
Takeaway:Run a leak check (WebRTC, DNS, system language, timezone) before each login, avoid web-page virtual-card billing, and leave quota headroom on new accounts instead of draining them within days.Related guide: Claude Anti-Ban Essentials: Quick Safety Checklist →
View original on X (Twitter) ↗Opens external link in a new tabSecond Claude ban — likely triggered by virtual U-card payment
Claude account banned for the second time. After ruling out IP and registration issues, the most likely cause is paying with a high-risk virtual U-card. Recommend using Wise inbound transfers or a real debit card instead.
Takeaway:Avoid high-risk virtual-card BINs; prefer Wise overseas accounts or real Visa/Mastercard debit cards.Related guide: Claude Payment Methods, Card Risk & Decline Troubleshooting →
View original on X (Twitter) ↗Opens external link in a new tabMulti-account failover strategy with isolated IPs
Production should never depend on a single Claude account. I run isolated residential IPs for separate Claude and ChatGPT backup accounts so a primary ban does not halt delivery.
Takeaway:Never single-point your production stack — build an isolated primary/backup account matrix.Related guide: Multi-Account Management Best Practices →
View original on X (Twitter) ↗Opens external link in a new tabClaude Code frequency and token-consumption risk boundaries
Documenting where high-frequency tool_use and automated requests cross from normal usage into model refusals or org-level disable. Short bursts of extreme concurrency are worse than sustained moderate load.
Takeaway:Throttle Claude Code automation frequency and concurrent tool calls — avoid burst patterns that look like distillation.Related guide: Safe Automation & Batch API Usage Guidelines →
View original on X (Twitter) ↗Opens external link in a new tabStable Claude Code setup amid recent ban wave
Sharing what has kept my Claude Code account stable through the recent ban wave: use a clean Gmail registered months ago, bind a real US phone number, and set an aged Gmail as recovery email.
Takeaway:Registration environment and email reputation matter — avoid disposable numbers and fresh throwaway inboxes.Related guide: Safe Account Registration, Virtual Card BIN & Payment Risk Avoidance →
View original on X (Twitter) ↗Opens external link in a new tabSuccessful reinstatement after official appeal — sharing confirmation email
Submitted an appeal through Anthropic's official channel with proof of legitimate academic/development use. Received reinstatement confirmation email within a few days.
Takeaway:Include verifiable academic or development-use proof in appeals; it gives reviewers something to check.Related guide: Account Ban Diagnosis & English Appeal SOP →
View original on X (Twitter) ↗Opens external link in a new tabAccount restored on day 2 via official appeal
Claude account disabled — submitted the standard English appeal through the official channel and got access back on day 2. Do not spin up replacement accounts while waiting.
Takeaway:After a ban, use the standard English appeal flow at claude.ai/restricted — reinstatement is achievable.Related guide: Account Ban Diagnosis & English Appeal SOP →
View original on X (Twitter) ↗Opens external link in a new tabClaude Code global deployment anti-ban guide
Claude Code deployment checklist: clean US dual-ISP static residential IP, OS timezone aligned with IP region, WebRTC leak prevention on Windows/Mac. Datacenter IPs are the #1 ban trigger.
Takeaway:Datacenter IPs are high-risk — pair Claude Code with residential broadband and fully aligned local timezone.Related guide: Claude Code Safe Configuration & Relay API Best Practices →
View original on X (Twitter) ↗Opens external link in a new tab