Anthropic releases Claude 4.5 with 1M context window
30-second read
Anthropic ships Claude 4.5 with a 1M token context, beating GPT-5 on coding benchmarks and matching Gemini 3 on multimodal. The thread is unusually concrete: real users posting real prompt regressions, real costs, and real opinions on whether the long-context era actually delivers.
▲ For
- Coding benchmarks (SWE-bench Verified, Aider polyglot) put Claude 4.5 ahead of every other frontier model — and the gap is largest on multi-file refactors, which matches user reports.
- Long context is genuinely useful for whole-codebase Q&A; multiple users report dropping their RAG pipeline for projects under 800k tokens.
- Pricing came down ~30% per 1M tokens at the same time, which (per the thread) is the biggest single quality-of-life win for heavy API users.
▼ Against
- Effective context degrades sharply past ~200k tokens — needle-in-haystack scores hold, but reasoning over the haystack does not. RAG is not dead.
- Long prompts are slow and expensive; one user posted a $14 single-shot bill for a 700k-token codebase Q&A.
- Prompt-regression reports: several users find Claude 4.5 worse than 4.0 on tightly structured outputs (e.g. SQL, JSON with strict schemas) — the model 'wants to explain itself' more than its predecessor.
? Still unknown
- Whether the 1M context is a true architectural win or mostly an inference-side trick (sliding window + retrieval) — Anthropic's paper is vague on details.
- How quickly OpenAI and Google match the price drop, and whether that triggers a real bottom-of-the-market race in 2026.
- Real-world tool-use reliability at long context — early benchmarks are encouraging but no large agent shop has shared production numbers yet.
→ What it means
- If you're building a coding agent, Claude 4.5 is the new default to beat — but evaluate yourself before switching, because long-context reasoning falls off a cliff sooner than the marketing suggests.
- RAG isn't going anywhere for production workloads above ~300k tokens or with strict latency SLAs.
- Pricing pressure is the real story — expect every frontier vendor to cut headline rates within the next quarter.
“ Notable quotes
The price drop is the headline. 1M context is a bonus. We've been waiting for serious competition at the GPT-4 price point for a year.— simonw
Benchmarks great, our internal eval mixed. It's better at writing code, worse at following our schema. Your mileage will vary, run your own evals.— ml_eng_anon
This is a hand-curated demonstration of grephn's AI compression format. Argument structure, quotes, and figures are representative of the linked HN discussion but are not verbatim transcriptions. Open the app to run live compression on any current thread.