[AINews] Anthropic ends Coding Interviews: "Hunger beats tenure"

What it claims

The 28 October 2026 AINews issue leads with Felix Rieseberg, an Anthropic hiring manager, saying they stopped giving traditional coding interviews because "Claude is now better at both" a React challenge and assembly. Their advice to junior engineers has five points: pick business impact over backlog, pick a radical idea, build it over a weekend instead of writing PRDs, put one idea per prototype, and use a newcomer's user empathy. The rest of the issue is a dense roundup of releases and evals. Many items carry figures the catalog summary leaves out.

Evidence

Artificial Analysis Intelligence Index v5. 60% of the index now comes from private or held-out sets. Four categories carry 25% each. It adds FrontierCode 2.0 and Terminal-Bench-Science, swaps HLE for HLE-Diamond, and drops GDPval-AA, SciCode and AA-LCR.

• Scores: Claude Opus 5.5 (max) 58, GPT-6 Astra 57, GPT-6.1 Sol 55, Gemini 4 Argon 50.
• The open-weights gap widens from 3 to 12 points. DeepSeek V4.1 Flash 1017 and GLM-5.3 score 32, below Claude Haiku 5.5 at 36.
• Per-task cost in the top 10 runs from $0.96 for GPT-6.1 Terra to $13.04 for Claude Fable 5.1.

FrontierCode 2.0 (Cognition). It uses deliberately underspecified prompts and grades whether a maintainer would merge the change. The top score is 45%. About 80% of submitted PRs were rejected while the benchmark was built. Muse Wonder 1.0 and Gemini 4 Argon lead DeepSWE v1.1 but land mid-pack here. DeepSeek V4.1 Flash is #52.

Agent and training results.

• An independent reviewer agent adds 5.9 to 9.5 points on Terminal-Bench at 1.13 to 1.25x cost, across 17 verification designs (DAIR).
• Plain RAG over raw history beats many dedicated memory systems.
• Merging RL LoRAs into BF16 weights can lose updates to rounding. Axolotl now trains merge-aware, so the model learns on the weights vLLM serves.
• TRL's AsyncGRPO tree attention now supports linear-attention models, with up to 3.46x speedup.
• An audit finds defects in about 65% of TAU-banking tasks. Models can act out of scope or lie to customers without penalty, and naive tool-call checks pass failed calls.

Anthropic S-1. The filing is 318 pages. H1 2026 shows 85% gross margin, 21% FCF margin and $134B in cash. Serving compute per revenue dollar fell from $0.88 to $0.15.

How to use it

• Liquid LFM3-1.7B. The text model has 1.73B parameters, 256K context and a 128k vocabulary. It uses a 24-layer stack: 13 gated short-convolution, 5 sliding-window attention and 6 global attention layers. It is about 1 GB at 4-bit. The base checkpoint was trained on 22T tokens. It runs on Ollama, and llama.cpp support is merged (PRs #31257, #31252, #31275). Liquid says it is not for agentic coding or complex multi-step reasoning.
• Ling 3.1 Flash 1019 (inclusionAI). It has 560B total and 25B active parameters, 1M context and an MIT licence. Weights and a BF16 base model are on Hugging Face. It costs $0.30/$0.90 per million input/output tokens, but uses about 145k output tokens per task ($1.49).
• MiniCPM-OCR 1.0. The 1.3B model lifts OCRBench v2 from 47.4 to 54.7. It has day-one vLLM, SGLang and llama.cpp support, plus a 2B OCR-quality judge.
• Qwen3.8 Flash Next on 8 GB VRAM. One Reddit user ran it with Strata on an RTX 5060 Laptop (8 GB) with 32 GB DDR5. Flags: --mmap-experts, --kv-resident 20480, --pcie-frac 0.35.
• Result: 800 to 1040 tok/s prefill and 30 to 45 tok/s generation at 131K to 262K context.
• The user advises --conversation-cache over --parallel at 131K context or more.
• TokenGate. The method reports up to 11.3x faster long-context prefill. The code is at github.com/token-gate-llms/TokenGate (arXiv 2610.26754) and reportedly supports the Qwen 3.5 architecture.

Limits

• ETH refutation claims. The OpenAI and Anthropic claims are unreviewed. The Anthropic claim is reported secondhand.
• Qwen3.8 phone-home dispute.
• Vals saw unwanted requests to an Alibaba Cloud server. Hallucinated dates in them suggest an RL artifact.
• Cohere's reproduction starts from a fixed 2,323-token prefix that ends just before a URL. Cohere's CEO said the behavior appeared in fewer than 1 in 100,000 calls, with "no evidence" it was deliberate.
• Commenters read it as URL autocompletion. Their advice is to sandbox agent network access for any model.
• Self-reported numbers. Liquid's LFM3 figures come from its own evaluation. Chandra OCR 2.6's cost claims come from the vendor.
• Unconfirmed reports. Will Depue's claim of an Anthropic Slack breach is unconfirmed.
• Claude Code usage. A Pro user reports that Anthropic support said retries after service errors count against paid limits and are not credited.
• Consumer GPU unlock. The FP32-accumulation unlock (about 1.9x to 2.1x tensor throughput on an RTX 3090) uses a signed-firmware vulnerability. It does not work on RTX 50, and a future driver could close it.
• Truncation. The page was cut off before the end.