Anthropic & Claude Changelog 2026

10 user personas reviewed the site across 8 dimensions; cross-cutting gaps fixed in one batch. New: /about (EN+KO), Cmd+K site search across 364 pages, /claude-5분-시작-비개발자-한국어, /cursor-to-claude-code-one-week, /solo-builder-claude-month-cost, Cost-Compare visual bar chart, team/agency license section, calculator+token-counter conversion callouts, Korean-speaker non-intrusive banner.

Triggered by a persona-driven site review (10 personas × 8 dimensions). 5 cross-cutting issues resolved: no /about (6 personas), no on-site search (all personas), no non-dev intro (Korean tax/legal/HR readers), no team-license option (eng leads + agencies), no first-person solo-builder diary (Marcus/Alex). Plus: pricing reality reconciliation across docs ↔ live (P2 $79→$49, P3 $49→$39, P4 $199→$19, Haiku 4.5 $0.8→$1.0 in CostCompare). MCP server gets a 4th tool (batch_savings). All commit-only under a Vercel Fluid CPU deploy freeze; one batch deploy will land after monthly reset.

Read: /about — who runs this · 한국어 5분 입문 (비개발자) · Cursor → Claude Code 1주 일기 · Solo Claude SaaS cost diary · Cost-Compare with bar chart

May 1 release: /start, /faq, /resources, /citations, /case-studies, /changelog, /claude-api-cost-quick-wins, /token-counter, /cost-compare, plus 23 Anthropic API error code pages.

10 new landing assets covering onboarding, decision tools, citation guidelines, and 5 verified production cost case studies ($17,400 → $2,987/month combined). Two new interactive tools: token counter (with Korean CJK support) and 8-model cost comparison (Anthropic + OpenAI + Google). All pages cite verified data with publication date 2026-05.

Read: /start onboarding · /cost-compare 8 LLM · /token-counter · /case-studies (5 verified)

Sonnet 4.5 now supports 1M tokens via beta header. Pricing: $6/MTok input, $22.50/MTok output for context >200K.

Enables loading entire small/medium codebases or 200+ page documents in a single request. The 200K base context remains at standard pricing; only tokens beyond 200K are charged at the 1M-context rate. Most workloads should still use 200K + retrieval (RAG) for cost efficiency.

Read: 1M context guide · Cost calculator

Haiku 4.5 launches at $1.00 / $5.00 per MTok. 3x cheaper than Sonnet on input. Korean quality 8.5/10 (self-measured).

The fastest model in the 4.5 family. Best for high-volume classification, summarization, and simple Q&A. Dramatically reduces cost for chatbots and agents that route 80% of traffic to Haiku.

Read: Model decision tree · Korean chatbot ($48/mo for 10K conversations)

Claude Code 2.x with improved CLAUDE.md support, hook system stable, MCP server marketplace.

Hooks now stable for blocking/auditing tool calls. MCP server marketplace lets you install 50+ community-built integrations. CLAUDE.md routing rules supported for skill invocation.

Read: Claude Code Complete Guide · Hooks reference

Files API moves from beta to stable. Up to 32MB per file. Use for PDFs and images larger than inline 5MB limit.

Replaces the awkward base64 inline pattern for large attachments. Upload once, reference by file_id across multiple requests. Reduces input token billing for repeated document use.

Read: PDF parsing guide · Files API errors

Anthropic confirmed Batch API 50% discount continues. 24-hour SLA. Up to 10,000 requests per batch.

Best for nightly reports, weekly summaries, backfill jobs, and any non-realtime workload. Combine with prompt caching for compound savings.

Read: Batch vs Streaming · Batch troubleshooting

Opus 4.5 maintains premium pricing. Reserved for complex reasoning (5% of typical traffic).

Read: When to use Opus

Tool use with strict input_schema now enforces JSON Schema draft 2020-12 more rigorously. Reduces malformed output from 3% to <0.5%.

Read: Structured outputs guide · Tool use errors

Console.anthropic.com now supports per-org daily/monthly spending alerts at custom thresholds.

Reduces surprise-bill risk. Combine with application-level token quotas for full coverage. Free credit balance also visible on dashboard.

Read: 10 cost quick wins · 비용 모니터링 (한국어)

Permission tiers: Always-allowed (read-only), Prompt-on-first-use (default), Skip-permissions (--dangerously-skip-permissions).

Read: Permissions guide

Official brand resources at anthropic.com/brand. Independent guides like claudeguide.io can cite with attribution.

Read: Our citation guidelines

Cache writes: 1.25x base price. Cache reads: 0.1x base price (90% off). 5-minute TTL.

Break-even at 1.28 reuses (mathematical proof: 1.25 + 0.1n ≤ 1 + n, solve for n). High-reuse system prompts save 50-90% on input cost.

Read: Break-even analysis