Skip to content
Source · Daily Brief

AI Daily Brief — 5 February 2026

A frontier-launch Thursday. Anthropic ships Claude Opus 4.6 — its first Opus-class model with 1M-token context (beta), new “agent teams”, context compaction, and outputs up to 128k tokens — at a 67% price cut to $5/$25 per M tokens. Opus 4.6 tops Terminal-Bench 2.0, leads Humanity’s Last Exam, scores 1606 Elo on GDPval-AA (a 144-point lead over GPT-5.2), and hits 60.7% on the Finance Agent benchmark. OpenAI counter-launches GPT-5.3-Codex — its first model classified “High” capability for cybersecurity under the Preparedness Framework, a new SOTA on SWE-Bench Pro and Terminal-Bench 2.0, and ~25% faster than 5.2-Codex. Two notable funding rounds — Fundamental ($1.4B unicorn for tabular AI) and Goodfire ($150M Series B at $1.25B for interpretability) — close on the same day.

Top stories

  • Claude Opus 4.6 launches. First Opus-class with 1M-token context (beta). New “agent teams” let multiple Claude sub-agents split larger tasks; context compaction auto-summarises older turns; outputs run up to 128k tokens. 67% price cut vs prior Opus — now $5 / $25 per M input/output tokens. Day-zero availability on the API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry. via Anthropic
  • Opus 4.6 benchmark sweep. Highest Terminal-Bench 2.0; leads Humanity’s Last Exam; 1606 Elo on GDPval-AA vs GPT-5.2’s 1462 and Opus 4.5’s 1416. Finance Agent: 60.7% (GPT-5.2 56.6%, Opus 4.5 55.9%, Sonnet 4.5 54.2%, Gemini 3 Pro 44.1%). 8-needle 1M MRCR v2 long-context retrieval: 76% vs Sonnet 4.5’s 18.5%. via Vellum
  • OpenAI — GPT-5.3-Codex launches. Unified Codex + GPT-5 stack; OpenAI’s “most capable agentic coding model” and 25% faster than GPT-5.2-Codex. New SOTA on SWE-Bench Pro and Terminal-Bench 2.0. Available on paid ChatGPT plans, with API access to follow. Sam Altman says the Codex team used early versions to debug its own training. via OpenAI
  • First “High” cyber rating. GPT-5.3-Codex is the first OpenAI model classified “High” for cybersecurity capability under the Preparedness Framework. The accompanying System Card details OpenAI’s “most comprehensive cybersecurity safety stack to date” — safety training, automated monitoring, trusted-access controls for advanced capabilities, plus a threat-intel-driven enforcement pipeline. Fortune leads on the framing: this is the first time OpenAI believes a model is good enough at coding and reasoning to meaningfully enable real-world cyber harm at scale. via OpenAI System Card
  • Microsoft Azure Foundry day-zero. Microsoft confirms Claude Opus 4.6 ships in Foundry on Azure at launch, joining Bedrock and Vertex AI for a true multi-cloud day-zero release — explicitly positioned for enterprise coding, agents and long-running workflows leveraging 1M-token context. via Microsoft

Money & infra

Fundamental exits stealth as a $1.4B unicorn with $255M total funding ($30M seed + $225M Series A). The Series A is led by Oak HC/FT with Valor Equity Partners, Battery Ventures, Salesforce Ventures and Hetz Ventures. Angel backers include Aravind Srinivas (Perplexity), Assaf Rappaport (Wiz), Henrique Dubugras (Brex) and Olivier Pomel (Datadog). The product: Nexus, a Large Tabular Model optimised for structured enterprise data. Goodfire closes a $150M Series B at a $1.25B valuation — less than a year after its Series A — led by B Capital with DFJ Growth, Salesforce Ventures and Eric Schmidt joining existing investors. Recent work: identifying a novel class of Alzheimer’s biomarkers via interpretability of an epigenetic foundation model (with Prima Mente), and halving LLM hallucinations via interpretability-informed training. via TechCrunch

Quiet corners

No first-party shipping from Google DeepMind, Meta AI, xAI, or Thinking Machines Lab. The Chinese pre-Lunar-New-Year hold continues. The day’s two frontier ships are unambiguously the lead story; Anthropic reclaims the professional-knowledge-work frontier (GDPval-AA, Finance Agent) while OpenAI raises the cyber-capability ceiling. The market splits on which strategy ages better.

By the numbers

  • Opus 4.6: 1M tokens (beta) · 128k output · $5/$25 per M (67% cut) · day-zero on Bedrock / Vertex / Foundry
  • Opus 4.6 benchmarks: Terminal-Bench 2.0 #1 · GDPval-AA +144 Elo vs GPT-5.2 · Finance Agent 60.7% · 1M MRCR v2 76%
  • GPT-5.3-Codex: “High” cyber rating (OpenAI first) · ~25% faster than 5.2-Codex · new SOTA SWE-Bench Pro + Terminal-Bench 2.0
  • Fundamental: $1.4B · $255M total ($30M seed + $225M Series A)
  • Goodfire: $150M Series B · $1.25B valuation

Compiled by AI Feed’s editor from publicly reported announcements on 5 February 2026.