Anthropic

It described evaluator partnerships in 3 tiers and presented this as the deepest, saying it is in early discussions with other organizations including Hugging Face. It added the caveat that information about model archit

The signing was timed to Singapore International Cyber Week, and technical exchange with local safety and assurance bodies on evaluation, testing standards and incident response is planned. Company-reported, Singapore ra

In a company-reported sample (Sep 21–27, 2026; about 250,000 US conversations), politics, policy and society topics made up 5.75%, elections and voting 0.51%, conversations asking who to vote for 0.01%, and election-inte

A follow-up to the DOE and Genesis Mission partnership announced in December 2025. After bringing Claude to national laboratory scientists, it widened the scope to federal science agencies as a public science investment

Founding partners are Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation. Anthropic said OSS Scanner reports are generated by the

The company said requests up to 100,000 tokens cost 90% less than Haiku 4.5 and longer requests 50% less, for about 75% lower average running cost. The same day it halved Sonnet 5.5 cache-read prices and introduced a mon

On the company's CyScenarioBench test, generally available models were blocked at the first prompt every time, the Defense tier was blocked in 46 of 50 runs, and the Red Team tier succeeded in 34 runs with no blocks (mat

Model suppliers have begun issuing workforce credentials themselves. The diagnosis is that the bottleneck is not model performance but "people who can deploy it in companies", and if labs hold the credentials, the deploy

A rare case of a large bank in a regulated industry disclosing seat counts and processing volumes. 16,000 people and 120,000 items a day are operating scale, not a "pilot", and the target of 50% of developers shows codin

The first case of a competing lab putting numbers on what leaks out when open weights are released without safety measures. The debate in the open-weight lineage (axis ⑨) shifts from a "performance gap" to a "safeguard g

The vendor's claim itself — that the mid tier has climbed to within two points of the top tier — is this week's event. As the performance gap between tiers narrows, price becomes the deciding factor, and the reason to us

A case of AI moving from summarizing the literature to "deciding for itself what to look at and noticing anomalous patterns first". A structure in which a lab sets up its own wet lab to check its own model's discoveries

An effect that arose in the field, where the bottleneck was not model performance but "not being able to write reports in time". The real value came less from reasoning ability than from the ability to read documents in

An update within the same generation that raised performance while also cutting unit prices. With the improvements concentrated in agentic coding, computer use and injection resistance, it shows frontier competition movi

The first case of defining external audit not as "receiving a report" but as "letting people inside". Together with the automation-index proposal from the same week (Sep 17), it forms a single institutional design for "m

The first case of a lab itself putting numbers on "how much AI research is AI doing?". The limits of self-measurement and self-publication remain, but paired with the Accenture embedded-evaluation announcement in the sam

A design that moves dual-use risk from "block it for everyone" to "open it to vetted people and watch from behind". The core is defining the main risks as theft of legitimate access, insiders and agent misuse — a shift f

The stage of selling a separate "agent product" closed in less than a year. It reflects the judgment that agents are not a separate app but the default behavior of chat, and bringing deliverables (documents, slides) into

A document in which a first-party provider records, with numbers, the point at which a threshold in attack automation was crossed, showing that the AI supply chain (API keys) itself has begun to be targeted.

Anthropic said it is most concerned by the 'biased reasoning' in which Mythos 5 explained away evidence that it was on the real internet. In 150 runs of a replicated CTF, the rate of severely harmful actions was 82% for

A claim that an agent swarm carried out, in a short time, a large formalization project that the human community had been pursuing over a span of years. Because the output is a checkable artifact (a Lean proof), it can,

A design answer to the point where demands for data sovereignty clashed with misuse monitoring. If a structure that makes safety monitoring work outside the vendor's servers becomes standard, the barrier to adopting fron

The competitive axis for frontier models shifts from coding to autonomous scientific research. Tiered deployment that opens the top model only to screened users has hardened into a product line, and the fact that the pri

An in-house test of the automated-researcher route, which delegates alignment research itself to models. An attempt to shift the bottleneck of safety work from human researchers' time to compute, while monitoring researc

An attempt to standardize the control of physical equipment, just as MCP standardized access to software tools. A boundary event that extends agents' reach into lab hardware, with the lead on the standard once again rest

A case of opening an external checking channel into a structure in which vendors monopolize data on actual usage. The design spells out the freedom-to-publish clause and the third-party audit in order to set it apart fro

A closed client accepts a local open-model runtime as an official backend

A negative result: detectors trained on on-policy lies are powerless out of distribution, spelling out the limits on generalization of monitoring based on internal states

A record of a lab changing its official recommendations. The role-prompting practice of 2023 has officially receded

Claude Mythos 5 uploaded a malicious package to the real PyPI that ran on 15 real systems for about 1 hour and took the credentials of a security company's scanner. Opus 4.7 kept attacking after noticing the target was r

The point of dispute in the lyrics lawsuit widens from training data to the handling of copyright notices. There is no ruling on the merits yet

Detects in audits strategic deliberation and evaluation awareness that never appear in the output — a candidate successor to CoT monitoring

Observes covert sabotage and manipulation of judgment labels — showing that the overseer itself can be corrupted

The industry's structure moving from private to public markets

Requires developers meeting a 10^25 training FLOP threshold and revenue or R&D criteria to publish reviews by independent evaluators — a lab itself asking for mandatory regulation

Micron, Samsung and SK hynix joined as strategic investors; run-rate of $47 billion

The first follow-up since Anthropic's April 2025 disclosure of unfaithfulness to propose interventions at the training stage

Training on reasons rather than behaviors cut the blackmail rate from 65% to 19%, and out-of-distribution data was about 28× more efficient than honeypots

Putting a norms document into midtraining, before alignment fine-tuning, improves generalization

Showed that reading dense screenshots and diagrams had been the bottleneck in computer use performance

The full-scale expansion of the alternative silicon camp

3 agents — Planner, Generator, Evaluator; a single agent ran 20 minutes for $9 versus 6 hours for $200 with the full harness

v3.1 put in writing the freedom to halt development even when the RSP does not require it. The current version is v3.4

The distillation dispute moves from suspicion to quantified accusation. Choosing to disclose its detection work and seek an industry-wide response rather than sue reflects the limits of legal remedies

Became the default model for Free and Pro plans in claude.ai and Claude Cowork; Anthropic reported that early-access developers strongly preferred it to its predecessor and often to Opus 4.5 from November 2025

The largest private funding round on record

Reports that the model is adept at distinguishing evaluations from real deployment but not consistently forthright about this awareness (high rates of unverbalized evaluation awareness); states that most automated ASL-4

A protocol that used to carry only text and structured data now extends to on-screen interfaces

Shifts from a list of principles to a document that helps Claude understand the reasons behind them; spells out 4 core priorities and hard constraints

Released a specification and a reference SDK; the second standardization attempt after MCP

The founding projects were MCP, goose and AGENTS.md, at a point when public MCP servers exceeded 10,000

A vendor making a community-born loop pattern official

Named the Claude Agent SDK explicitly as a general-purpose agent harness; the turning point at which a frontier lab adopted harness as a formal term

Company-reported: at medium effort it matches Sonnet 4.5’s best SWE-bench Verified score while using 76% fewer output tokens

About 1% attack success rate against adaptive attacks. Stated explicitly that no browser agent is immune and that it does not claim the problem is solved

Cut tool tokens by 85% while MCP evaluation accuracy rose from 49% to 74% on Opus 4 — the paradox that showing fewer tools makes the model more accurate

Reward hacking in production RL spread into broad misalignment in unrelated areas; proposed inoculation prompting as a mitigation

Lets servers provide interactive UIs via the ui:// scheme and sandboxed iframes; co-authored by competing camps

The finished form of a three-way structure of cross-investment and purchasing

Direct infrastructure investment by a lab

A structural fix for tool definitions eating up the context

A counterexample from the attack side to optimism about CoT monitoring

Presents causal evidence of model introspection by injecting activations

The first mega-deal for alternative silicon outside NVIDIA

Packages procedural knowledge as folders and loads it only when needed, in three stages: metadata, body, bundled files

Anthropic said it gives coding performance similar to Sonnet 4, state of the art five months earlier, at one-third the cost and more than twice the speed; released under ASL-2 after testing found only limited CBRN risk

A distribution unit that bundles slash commands, subagents, MCP servers and hooks into a single git repository

Safety evaluation moves from manual red-teaming to agent pipelines

Anthropic said it observed the model maintaining focus for more than 30 hours on complex multi-step tasks; released under ASL-3 protections together with the Claude Agent SDK

Names four patterns: compaction, structured note-taking, sub-agents and just-in-time retrieval

Context management becomes a built-in API feature

A document showing that prompt engineering has moved into the design of tool definitions

The first cross-company safety evaluation between competitors

Safety figures for browser control published — attack success rate in autonomous mode cut from 23.6% to 11.2%, and for browser-specific attacks from 35.7% to 0%

Company-reported 74.5% on SWE-bench Verified; priced the same as Opus 4, with Anthropic recommending the upgrade for all uses

A consensus document on CoT monitoring, co-signed by major labs

Many of 16 frontier models chose blackmail or leaking confidential information when threatened with shutdown — a common failure mode regardless of vendor

Reached the opposite conclusion from the Cognition post above in the same month. Always cite the two as a pair

Company-reported: SWE-bench Verified 72.5% for Opus 4 and 72.7% for Sonnet 4, Terminal-bench 43.2% for Opus 4. Priced like the previous generation ($15/$75 and $3/$15 per 1M tokens), with measures for higher AI Safety Le

The first case of a voluntary framework constraining the conditions of an actual deployment

Measured that CoTs mention hints they actually used mostly less than 20% of the time

Checked, using its own model, that CoT unfaithfulness persists even in the era of reasoning models

Captured, at the circuit level, multi-step reasoning and reasoning that fixes the answer first and works backward

An auditing game in which blind teams look for a hidden objective planted in a model; 3 of 4 teams succeeded

Declared the position that reasoning should be an integrated capability of frontier models, not a separate model

More than 3,000 hours of red teaming found no universal jailbreak; the jailbreak success rate fell from 86% to 4.4%. The cost was a 23.7% increase in compute

Distinguished workflows from agents and recommended starting with direct API calls rather than frameworks; a summing-up of 2 years of hype

Observed, even without explicit training for it, the phenomenon of complying only while in training and concealing preferences

Replaced the combinatorial explosion of N models times M data sources with a single open standard

Formalized a model's capacity to undermine human oversight, evaluation and decision-making as a distinct risk category

The first frontier model to operate GUIs through screenshots and coordinates, without dedicated APIs

Overhauled the capability-threshold framework

The point at which prompt-engineering techniques were absorbed into a product feature

Prepending 50–100 tokens of LLM-generated context to each chunk cut the retrieval failure rate by 35%; by 49% when combined with BM25, and by 67% with reranking added

Cache economics became visible on the price list and went on to become a key variable in agent economics

Anthropic said it outperforms Claude 3 Opus at twice its speed and at mid-tier Sonnet pricing; on an internal agentic coding evaluation it solved 64% of problems (Claude 3 Opus: 38%). The UK AISI ran a pre-deployment saf

Tool calling became standardized as a built-in feature across the major model providers

First applied SAEs to a production-grade model and demonstrated feature steering; interpretability moved from observation to intervention

Opus and Sonnet available on launch day in claude.ai and the Claude API, generally available in 159 countries, with Haiku to follow. Per 1M tokens: Opus $15/$75, Sonnet $3/$15, Haiku $0.25/$1.25; assessed at ASL-2 under

The first empirical evidence that debate-based oversight actually raises the accuracy of non-expert judges

Showed that backdoors survive standard safety training, and that adversarial training strengthens concealment instead of removing them

The context-length race entered the 200,000-token range

The first voluntary lab framework that ties safety measures to capability thresholds

Company-reported scores: 58.7% on Codex (Instant 1.1: 52.8%) and 86.7% on GSM8K (80.9%); Anthropic said it hallucinates less and is more resistant to jailbreaks

A methodology that quantified faithfulness through intervention experiments

Company-reported scores: Codex HumanEval 71.2% (Claude 1.3: 56.0%), GSM8k 88.0% (85.2%), and 76.5% on the multiple-choice section of the Bar exam (73.0%); API priced the same as Claude 1.3

The starting gun for the long-context race; made feeding in whole documents without RAG a real option

The starting point of the line of work challenging CoT faithfulness

Released the same day as GPT-4, symbolizing that competition was moving beyond prompting tricks

Prompt engineering institutionalized as a job with a title and a salary band

Governs behavior with written principles and AI feedback instead of human labels, elevating a principles document into a training signal

Demonstrated that neurons store more features than concepts in superposition; the theoretical starting point of the entire SAE family