Key Terms
- Token-based billing - charges based on the number of input and output tokens processed. One token is roughly 4 characters or 0.75 words. Source: Claude – Token Usage
- Prompt caching - stores frequently used prompt prefixes to cut latency and cost. 5-minute TTL cache reads cost 0.1x of standard input; a 1-hour extended TTL is also available. Source: Claude – Pricing
- Batch API - asynchronous processing that costs 50% less than standard API calls, with results within 24 hours. Source: Claude – Batch Processing
- Context window - the maximum tokens a model can process in one conversation. Opus 4.8, Sonnet 5, and Haiku 4.5 all support a 1M token context window. Source: Anthropic – Pricing
- Effort levels - controls how much reasoning a model performs before answering. Sonnet 5 ships with adaptive (thinking) mode on by default and can be set to
disabled. Source: Anthropic – Claude Sonnet 5 - Fast mode - a research preview giving up to 2.5x faster output at premium pricing. Opus 4.8 fast mode is $10/$50 per MTok (2x standard). Source: Claude – Pricing
- Mythos-class models - Anthropic's highest-capability tier (Fable 5, Mythos 5) aimed at advanced cyber and intelligence workloads. They carry a mandatory 30-day customer-data retention period and strict cyber safeguards. Source: Support – 15425996 Data Retention Practices For Mythos Class Models
- Cyber Verification Program - real-time safeguards that detect and block dangerous cyber usage. Enabled by default on Sonnet 5, Opus 4.7, and Opus 4.8. Source: Support – 14604842 Real Time Cyber Safeguards On Claude
- Agent SDK credit - a planned (now paused) monthly credit for programmatic usage (
claude -p, Agent SDK, third-party Agent SDK apps) billed at API rates instead of drawing from subscription limits. Source: News – Item
Latest Changes
Changes since the 2026-05 report.
- New model: Claude Sonnet 5 launched June 30 at introductory $2/$10 per MTok through August 31, 2026, then $3/$15 standard. Default model for Free and Pro; available in Claude Code and on the API (
claude-sonnet-5). Source: Anthropic – Claude Sonnet 5 - Tokenizer change: Sonnet 5 uses an updated tokenizer that produces roughly 30% more tokens than Sonnet 4.6 for the same input (1.42x English, 1.33x Spanish, 1.28x Python code, ~1.01x Simplified Chinese). Intro pricing is set to be "roughly cost-neutral" for the transition. Source: Simonwillison – Claude Sonnet 5
- API behavior change: Sonnet 5 drops support for the
temperature,top_p, andtop_ksampling parameters; adaptive thinking is on by default. Max output is 128,000 tokens over a 1M context window. Source: Simonwillison – Claude Sonnet 5 - Major event (governance): US government export-control directive (June 12) forced Anthropic to disable Fable 5 and Mythos 5 for all customers, citing a national-security jailbreak concern. Source: Anthropic – Fable Mythos Access
- Feature restored: Fable 5 cleared to return globally July 1, alongside a proposed industry-wide jailbreak-severity scoring framework co-developed with Amazon, Microsoft, and Google. Mythos 5 remains suspended. Source: Anthropic – Redeploying Fable 5
- Plan change (planned, then paused): The Agent SDK credit change scheduled for June 15 was paused by Anthropic on June 15.
claude -pand Agent SDK usage continues to draw from subscription limits as before; no credit to claim. New timing undisclosed. Source: News – Item - Product added: Claude Tag (@Claude) launched June 23 as a new way for teams to work with Claude. Source: Anthropic – Introducing Claude Tag
- Product added: Claude Science, an AI workbench for scientists, is now available (June 30). Source: Anthropic – Claude Science Ai Workbench
- Partnerships: TCS (regulated industries, June 12) and DXC (banks, airlines, regulated industries, June 11) alliances announced. Source: Anthropic – Tcs Anthropic Partnership
- Expansion: Seoul office opened June 17 with new Korean ecosystem partnerships. Source: Anthropic – Seoul Office Partnerships Korean Ai Ecosystem
- Policy: "Policy on the AI Exponential" published June 10, proposing institutional preparation for AI progress. Source: Anthropic – Policy On The Ai Exponential
- Rate limits (retroactive): On April 26 Anthropic raised Sonnet and Haiku rate limits at every usage tier and simplified to three tiers (Start, Build, Scale) on the native Platform. Source: Anthropic – Claude Sonnet 5
Plans
| Plan | Price (monthly) | Price (annual) | Usage vs Free | Key Inclusions |
|---|---|---|---|---|
| Free | $0 | $0 | Baseline | Web/iOS/Android/desktop chat, code execution (50 hrs/day per org), web search, memory, connectors, MCP, extended thinking, desktop extensions |
| Pro | $20/mo | $17/mo ($200/yr billed up front) | "More usage" (undisclosed) | Everything in Free + Claude Code, Claude Cowork, Claude Design, unlimited projects, Research, more models, Claude for Microsoft 365 and Outlook |
| Max 5x | $100/mo | undisclosed | 5x Pro usage (undisclosed) | Everything in Pro + higher output limits, early access, priority access at high traffic |
| Max 20x | $200/mo | undisclosed | 20x Pro usage (undisclosed) | Everything in Max 5x + maximum limits |
| Team Standard | $25/seat/mo | $20/seat/mo | More than Pro (undisclosed) | SSO, central billing, enterprise search, no model training on content, admin controls, mix/match seat types |
| Team Premium | $125/seat/mo | $100/seat/mo | 5x standard seat usage (undisclosed) | Everything in Team Standard + 5x usage |
| Enterprise | $20/seat + usage at API rates | annual | Usage scales with API spend | SCIM, audit logs, compliance API, HIPAA-ready, custom data retention, IP allowlisting, RBAC, spend controls, Claude Security (beta) |
Agent SDK credit: PAUSED on June 15. No credit is currently claimable; claude -p and Agent SDK usage continues to draw from subscription limits exactly as before. Source: News – Item
Source: Anthropic – Pricing
API Pricing
| Model | Input ($/MTok) | Output ($/MTok) | Cache Write 5m ($/MTok) | Cache Read ($/MTok) | Batch Discount |
|---|---|---|---|---|---|
| Fable 5 | $10.00 | $50.00 | $12.50 | $1.00 | 50% |
| Opus 4.8 | $5.00 | $25.00 | $6.25 | $0.50 | 50% |
| Opus 4.8 (fast mode) | $10.00 | $50.00 | $12.50 | $1.00 | N/A |
| Sonnet 5 (intro, through Aug 31 2026) | $2.00 | $10.00 | $2.50 | $0.20 | 50% |
| Sonnet 5 (standard, after Aug 31 2026) | $3.00 | $15.00 | $3.75 | $0.30 | 50% |
| Haiku 4.5 | $1.00 | $5.00 | $1.25 | $0.10 | 50% |
| Opus 4.7 (legacy) | $5.00 | $25.00 | $6.25 | $0.50 | 50% |
| Opus 4.6 (legacy) | $5.00 | $25.00 | $6.25 | $0.50 | 50% |
| Sonnet 4.6 (legacy) | $3.00 | $15.00 | $3.75 | $0.30 | 50% |
| Opus 4.1 (legacy) | $15.00 | $75.00 | $18.75 | $1.50 | 50% |
Status notes:
- Fable 5: suspended for all customers June 12 under US export-control directive; cleared to return July 1. Source: Anthropic – Redeploying Fable 5
- Mythos 5: suspended June 12; no restoration date announced. Source: Anthropic – Fable Mythos Access
Other API pricing:
- Managed Agents: $0.08 per session-hour for active runtime
- Web search: $10 per 1,000 searches (plus token costs)
- Code execution: 50 free hours per day per organization, then $0.05/hour per container
- US-only inference (data residency): 1.1x multiplier on input/output tokens
- Batch API: 50% discount on all models
- 1M token context window included at standard pricing for Opus 4.8, Sonnet 5, Haiku 4.5
Source: Anthropic – Pricing
Model Performance / Benchmarks
Sonnet 5 benchmark headline scores were published primarily as chart images in the launch blog; the full numbers are in the system card PDF. Verifiable text-based data points are below.
| Benchmark / Metric | Sonnet 5 | Sonnet 4.6 | Opus 4.8 (reference) | Notes |
|---|---|---|---|---|
| Overall positioning | "performance close to Opus 4.8, but at lower prices" | predecessor | reference frontier | Per Anthropic. Source: Anthropic – Claude Sonnet 5 |
| CyberGym (with default mitigations) | 0 | n/a | higher | Sonnet 5 "far less capable" at cyber than Opus 4.8 and Mythos 5. Source: News – Item |
| Firefox 147 exploit (working) | 0.0% | 0.0% | higher | Neither Sonnet model could build a working exploit; Sonnet 5 showed slightly higher partial success than 4.6. Source: Anthropic – Claude Sonnet 5 |
| OSWorld-Verified (computer use) | undisclosed (strict improvement over 4.6) | 78.5% (re-scored) | higher | Sonnet 4.6 score updated from 78.0% after methodology change. Source: Anthropic – Claude Sonnet 5 |
| Humanity's Last Exam (no tools / with tools) | undisclosed | 34.6% / 46.8% (re-scored) | higher | Grader model updated; Sonnet 4.6 figures revised. Source: Anthropic – Claude Sonnet 5 |
| Misaligned-behavior audit | lower (safer) than Sonnet 4.6 | baseline | lower than Sonnet 5 | Sonnet 5 safer than 4.6 overall, but higher misalignment rate than Mythos Preview and Opus 4.8. Source: Anthropic – Claude Sonnet 5 |
Note: Anthropic did not publish Sonnet 5's exact SWE-Bench Verified or BrowseComp headline numbers in extractable text form; they appear in chart images and the system card PDF. Community benchmarkers report Sonnet 5 at roughly "GLM-5.2 level" on independent coding/proofreading tests. Sources: Anthropic – Claude Sonnet 5 , News – Item
Latest News
Claude Sonnet 5 Launch (June 30, 2026)
Anthropic launched Claude Sonnet 5, the most agentic Sonnet yet, positioned as close to Opus 4.8 capability at lower prices. It is the default model for Free and Pro plans and is available in Claude Code and on the Claude Platform at introductory $2/$10 per MTok through August 31, then $3/$15. It ships with cyber safeguards on by default, adaptive thinking on by default, a 1M context window, 128K max output, and a new tokenizer that increases token counts by roughly 30% versus Sonnet 4.6. Source: Anthropic – Claude Sonnet 5
US Government Suspends Fable 5 and Mythos 5 (June 12, 2026)
The US government issued an export-control directive forcing Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national, including foreign-national Anthropic employees, effectively disabling both models for all customers. The directive cited a national-security jailbreak concern. Anthropic stated the capability level in question is widely available in other public models, including OpenAI's GPT-5.5, and is complying under protest. Source: Anthropic – Fable Mythos Access
Fable 5 Redeployed (June 30, 2026)
Fable 5 was cleared to return globally on July 1. Anthropic also proposed an industry-wide framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other Glasswing partners. Mythos 5 was not included in the redeployment. Source: Anthropic – Redeploying Fable 5
Agent SDK Credit Change Paused (June 15, 2026)
The Agent SDK credit change scheduled for June 15 was paused on the day it was due to take effect. Anthropic emailed subscribers that claude -p, the Agent SDK, and third-party Agent SDK apps continue to draw from subscription limits as before, with no credit to claim and no new date given. Source: News – Item
Claude Tag (June 23, 2026)
Anthropic launched Claude Tag (@Claude), described as a new way for teams to work with Claude. Source: Anthropic – Introducing Claude Tag
Claude Science (June 30, 2026)
Claude Science, a customizable AI workbench for scientists that integrates research tools and produces auditable artifacts, is now available. Source: Anthropic – Claude Science Ai Workbench
Enterprise Partnerships and Expansion (June 2026)
TCS (June 12) and DXC (June 11) announced alliances to deploy Claude in regulated industries (banking, airlines). Anthropic opened a Seoul office (June 17) and announced Korean ecosystem partnerships. The first "Anthropic Public Record" results were published (June 12). Sources: Anthropic – Tcs Anthropic Partnership , Anthropic – Dxc Anthropic Alliance , Anthropic – Seoul Office Partnerships Korean Ai Ecosystem
Policy on the AI Exponential (June 10, 2026)
Anthropic published policy proposals arguing that institutions must be prepared for exponential AI progress, including government authority to block unsafe deployments through a transparent, technically grounded process. Source: Anthropic – Policy On The Ai Exponential
Community Signals
Sonnet 5 Launch Thread: Mixed, with Sharp Price/Performance Critique
The HackerNews thread for the Sonnet 5 launch reached 1,099 points and 639 comments. The dominant theme was that, after accounting for the new tokenizer, Sonnet 5 is more expensive than its headline price implies, and several independent benchmarkers found it behind competing frontier models on price/performance. Source: News – Item
Key quotes:
- conradkay: "Seems worse even on price/performance than GLM 5.2, which is only 744b parameters. The system card says that with default mitigations Sonnet 5 scored a 0 on CyberGym, far less capable than Opus 4.8 and Mythos 5." News – Item
- XCSme: "On my benchmarks it's GLM-5.2 level at 2x the cost, but also 2x faster. Weak spots: Trivia 0/3 (little built-in knowledge), combined tool-calling 45/100 (sometimes makes invalid tool calls), and Puzzle Solving 77." News – Item
- artursapek: "On my proofreading benchmark, Sonnet 5 is definitely better than Sonnet 4.6, but inferior on both quality and cost to GLM 5.1, GLM 5.2, Gemini 3.1 Flash, and Gemini 3.1 Pro." News – Item
- kingjimmy: "The footnotes reveal Sonnet 5 uses an updated tokenizer where the same input can map to roughly 1.0-1.35x more tokens than Sonnet 4.6. AKA expect higher costs on Sonnet 5 for the same tasks." News – Item
Not all feedback was negative. A subset of developers reported strong agentic gains:
- boutell: "Powering our CMS editing agent, Sonnet 5 is a LOT better out of the box than Sonnet 4, one-shotting complex instructions that Sonnet 4 used to miss half of, and independently recovering from bad input that caused 400 errors." News – Item
- ashvardanian: "I swapped Opus subagents for Sonnets on 3 different projects and not one wrote a single line of code; they just kept spinning and wasting tokens. Can't remember the last time that happened with Opus. Reverting back." News – Item
Tokenizer "Stealth Price Hike" Narrative
Simon Willison's independent tokenizer tests quantified the token inflation at roughly 1.42x for English text, 1.33x for Spanish, and 1.28x for Python code, concluding the new tokenizer is "effectively a 30% price increase" once the intro pricing expires. Source: Simonwillison – Claude Sonnet 5
SDK Credit Pause: Relief, but Trust Erosion
The June 15 pause of the Agent SDK credit was covered across multiple HN submissions. The highest-engaged thread ("Anthropic pauses credit change for Claude Code", 36 points) framed the reversal as a rare walk-back, though commenters noted the underlying push toward API-rate billing for programmatic use is likely to return. Source: News – Item
Enterprise Readiness
| Feature | Available? | Details |
|---|---|---|
| SSO (SAML) | Yes | SAML 2.0 (Okta, Azure AD). Team plan and above. Source: Anthropic – Pricing |
| SSO (OIDC) | No | Only SAML 2.0 listed. OIDC not mentioned on pricing or enterprise pages. |
| SCIM | Yes | Enterprise plan only. Source: Anthropic – Pricing |
| Audit logs | Yes | Enterprise plan only. Source: Anthropic – Pricing |
| IP indemnity | No | Not mentioned on any pricing, enterprise, or product page. |
| Data residency | Partial | US-only inference available via API at 1.1x token multiplier. No EU or APAC-only option. Source: Claude – Data Residency |
| HIPAA | Yes | HIPAA-ready offering on Enterprise plan. Source: Anthropic – Pricing |
| Air-gapped / on-prem | No | No native air-gapped deployment. BYOK via Amazon Bedrock, Google Vertex AI, or Microsoft Foundry within your own VPC. Source: Claude – Enterprise |
| SLA | No | No publicly documented SLA on pricing or enterprise pages. |
| Admin controls (RBAC) | Yes | Role-based access, user/org spend controls, IP allowlisting, network-level access control. Enterprise plan. Source: Anthropic – Pricing |
| Claude Security | Yes (beta) | Enterprise plan. Source: Anthropic – Pricing |
Transparency Gaps
| Gap | Details | Severity |
|---|---|---|
| Mythos 5 availability | Suspended June 12 with no restoration date. Buyers relying on Mythos-class capability have no timeline. | High |
| Usage limits per plan | No exact token/request limits published for any tier. Even "5x Pro usage" and "20x Pro usage" have no concrete number. | High |
| Agent SDK credit timing | Paused June 15 with no replacement date. Buyers cannot plan for when programmatic usage will move to metered billing. | High |
| Sonnet 5 benchmark scores | Headline SWE-Bench Verified / BrowseComp numbers published only as chart images; full detail confined to the system card PDF. | Medium |
| Tokenizer cost impact | Anthropic frames the new tokenizer as "cost-neutral" at intro pricing, but after Aug 31 the 30% token inflation becomes a real price increase versus Sonnet 4.6. Not surfaced on the pricing page. | High |
| Effort/token multipliers | The token multiplier for higher effort levels vs high is not disclosed. Anthropic says higher effort "drives token consumption through the roof" without quantifying. |
Medium |
| Fast mode availability | Opus 4.8 fast mode is listed, but eligibility, rate limits, and capacity guarantees are not documented. | Low |
| Rate-limit tiers | The new Start/Build/Scale API rate-limit tiers (introduced April 26) are visible only in the Console, not published as concrete RPM/TPM numbers. | Medium |