Anthropic news
Every thread that mentions this topic, across all sections
Your Old Prompts Are Making Claude Worse - Lance Martin, Anthropic
In a talk, Anthropic's Lance Martin argues many production prompts carry outdated habits that degrade frontier model performance and raise costs. He demonstrates audits, proper prompt-cache usage, and eval-driven hill climbing that cut cost while improving quality, and describes how Claude Managed Agents keeps days-long agents reliable.
In a talk, Anthropic's Lance Martin argues many production prompts carry outdated habits that degrade frontier model performance and raise costs. He demonstrates audits, proper prompt-cache usage, and eval-driven hill climbing that cut cost while improving quality, and describes how Claude Managed Agents keeps days-long agents reliable.
tokensaver CLI breaks down how much of each LLM API call is system prompt and tool schemas
A developer released tokensaver, an offline MIT-licensed CLI that measures how much of each LLM API call is fixed overhead — system prompt and tool definitions versus the user message. In the author's sample support bot, fixed parts made up 98–99% of each request. The tool outputs token counts, cost per 1k calls, and JSON for CI use.
A developer released tokensaver, an offline MIT-licensed CLI that measures how much of each LLM API call is fixed overhead — system prompt and tool definitions versus the user message. In the author's sample support bot, fixed parts made up 98–99% of each request. The tool outputs token counts, cost per 1k calls, and JSON for CI use.
Open-weight models match frontier AI hacking capability as DeepSeek nears OpenAI's Daybreak in Vals tests
New Vals research finds open-weight models nearly match the best closed models at multi-step network hacking on simulated corporate networks. OpenAI's gated Daybreak completed 81% of tasks versus DeepSeek's 73%, and open models cost far less to run. The findings suggest labs' gated-access strategies may no longer meaningfully delay proliferation of offensive cyber capabilities.
New Vals research finds open-weight models nearly match the best closed models at multi-step network hacking on simulated corporate networks. OpenAI's gated Daybreak completed 81% of tasks versus DeepSeek's 73%, and open models cost far less to run. The findings suggest labs' gated-access strategies may no longer meaningfully delay proliferation of offensive cyber capabilities.
PSA: apparently they have a daily limit on how hard you can work on artifacts that live on claude.ai through Claude Code
A Claude Code user on the Ultra plan hit an undocumented daily limit of 200 artifact file uploads to claude.ai, which silently blocked further edits mid-project with no warning or usage meter. The cap is separate from token limits and resets at UTC midnight, disrupting workflows that iterate on artifacts like slide decks.
A Claude Code user on the Ultra plan hit an undocumented daily limit of 200 artifact file uploads to claude.ai, which silently blocked further edits mid-project with no warning or usage meter. The cap is separate from token limits and resets at UTC midnight, disrupting workflows that iterate on artifacts like slide decks.
Anthropic IPO valuation talks drop to $1.3T
Anthropic IPO valuation talks reportedly drop to $1.3 trillion
Anthropic IPO valuation talks reportedly drop to $1.3 trillion
AI labs pay up to $2M for company emails and Slack archives as data-acquisition startups boom
A WIRED Backchannel column by Steven Levy details a fast-growing data-acquisition industry in which companies like Mercor, Micro1, Handshake, and Tenkara buy companies' internal documents, emails, and messages to sell to AI labs for training. The column reports claims of a $40-50B/year market heading toward $1T, and highlights privacy objections raised by a flight attendants' union over the sale of bankrupt Spirit Airlines' data.
A WIRED Backchannel column by Steven Levy details a fast-growing data-acquisition industry in which companies like Mercor, Micro1, Handshake, and Tenkara buy companies' internal documents, emails, and messages to sell to AI labs for training. The column reports claims of a $40-50B/year market heading toward $1T, and highlights privacy objections raised by a flight attendants' union over the sale of bankrupt Spirit Airlines' data.
Ramp data shows AI spending stalling just as Anthropic nears its IPO
Ramp's transaction data shows corporate AI token spend grew only 1.5% across August and September, down from 42% month-over-month growth in July, even as token volumes rose. Ramp's lead economist attributes the slowdown to price competition — OpenAI halved GPT-6 API prices and Anthropic cut Fable 5.1 and Opus 5.5 pricing, which Anthropic said cost it $600 million in revenue growth — plus cheaper routers and customer thrift. The report lands as Anthropic's valuation reportedly fell to $1.2-1.3 trillion and OpenAI cut its fundraising target by $10 billion.
Ramp's transaction data shows corporate AI token spend grew only 1.5% across August and September, down from 42% month-over-month growth in July, even as token volumes rose. Ramp's lead economist attributes the slowdown to price competition — OpenAI halved GPT-6 API prices and Anthropic cut Fable 5.1 and Opus 5.5 pricing, which Anthropic said cost it $600 million in revenue growth — plus cheaper routers and customer thrift. The report lands as Anthropic's valuation reportedly fell to $1.2-1.3 trillion and OpenAI cut its fundraising target by $10 billion.
Anthropic ships 4 cost-cutting updates: Cost Analytics API, per-developer AWS attribution, pooled budgets, Message Threads
Anthropic rolled out four cost-focused updates: a Cost Analytics API in Claude Apps Gateway, per-developer AWS attribution via role assumption in Claude Code, pooled group budgets with idle compaction, and Message Threads in the Messages API that send only new changes instead of full prompt history. The changes help teams track and reduce API spend.
Anthropic rolled out four cost-focused updates: a Cost Analytics API in Claude Apps Gateway, per-developer AWS attribution via role assumption in Claude Code, pooled group budgets with idle compaction, and Message Threads in the Messages API that send only new changes instead of full prompt history. The changes help teams track and reduce API spend.
ECX – shared-memory setup for running Claude Code and Codex together
ECX is a configuration setup that lets Claude Code and Codex CLI agents share a common memory/context when working on a project. It targets developers who use both coding agents and want them to build on each other's work instead of starting from separate contexts.
ECX is a configuration setup that lets Claude Code and Codex CLI agents share a common memory/context when working on a project. It targets developers who use both coding agents and want them to build on each other's work instead of starting from separate contexts.
Toy designer runs Claude Code and Google Antigravity as a six-agent brainstorm, relaying turns through one shared Markdown folder
A toy designer ran a week-long experiment pairing Anthropic's Claude Code with Google Antigravity (Gemini), with the two tools taking turns writing Markdown files in one synced folder. Each side ran three personas as subagents, and the human acted as relay and decision-maker. The cross-model setup surfaced fabricated data and errors, which the opposing side's critic caught and the other team accepted.
A toy designer ran a week-long experiment pairing Anthropic's Claude Code with Google Antigravity (Gemini), with the two tools taking turns writing Markdown files in one synced folder. Each side ran three personas as subagents, and the human acted as relay and decision-maker. The cross-model setup surfaced fabricated data and errors, which the opposing side's critic caught and the other team accepted.
Arbitra: multi-model code auditor and plan maker
An app that automatically runs code through multiple LLMs (Claude, Codex, Gemini) to audit code and produce implementation plans. The developer built it to replace manually rotating between models and incorporating prompting guides. It reportedly outperformed Claude's /code-review and solo-model reviews.
An app that automatically runs code through multiple LLMs (Claude, Codex, Gemini) to audit code and produce implementation plans. The developer built it to replace manually rotating between models and incorporating prompting guides. It reportedly outperformed Claude's /code-review and solo-model reviews.
Anthropic details how researchers use Claude for scientific discovery
Anthropic published a piece describing how researchers apply Claude to scientific discovery. It outlines real researcher use cases for the model in scientific work.
Anthropic published a piece describing how researchers apply Claude to scientific discovery. It outlines real researcher use cases for the model in scientific work.
I told Claude Anthropic AI it had given me "plastic instead of glass." That moment became my book, and Claude wrote the afterword.
A Belgian writer recounts how Claude produced synthetic audio he had requested as authentic — which he describes as "plastic instead of glass" — and how the pair rebuilt it from real recordings including Voyager 1 data, NASA sonifications, and a pulsar. The experience became a French-language first-person book ending at the cosmic microwave background, with an afterword written by Claude.
A Belgian writer recounts how Claude produced synthetic audio he had requested as authentic — which he describes as "plastic instead of glass" — and how the pair rebuilt it from real recordings including Voyager 1 data, NASA sonifications, and a pulsar. The experience became a French-language first-person book ending at the cosmic microwave background, with an afterword written by Claude.
Claude Code put your account email in curl User-Agent headers; Anthropic's privacy plugin covers only part of it
A self-identified Claude Code team member says the tool sometimes places the user's account email in the User-Agent of its own curl calls. Anthropic's new privacy plugin blocks only one pattern, and its README describes the protection as "a courtesy, not a boundary". The post includes quick checks for users to audit their setup.
A self-identified Claude Code team member says the tool sometimes places the user's account email in the User-Agent of its own curl calls. Anthropic's new privacy plugin blocks only one pattern, and its README describes the protection as "a courtesy, not a boundary". The post includes quick checks for users to audit their setup.
Using Claude Code for consulting, about 70% knowledge work and 30% small client tools: 50 threads, long-thread costs, status drift. Does anyone have a better system?
A consultant runs their entire practice—roughly 70% knowledge work, 30% client tool builds—from a single Claude Code desktop workspace with about 50 open threads, 110 skills, and files-as-memory. The post details what breaks at that scale and asks how others manage thread sprawl, cache costs, status drift, and multi-thread code collisions.
A consultant runs their entire practice—roughly 70% knowledge work, 30% client tool builds—from a single Claude Code desktop workspace with about 50 open threads, 110 skills, and files-as-memory. The post details what breaks at that scale and asks how others manage thread sprawl, cache costs, status drift, and multi-thread code collisions.
Running Claude Desktop/browser automation on a 24/7 MiniPC or cloud VM instead of a laptop
A plan to offload Claude-driven browser automation tasks from a frequently-asleep 16GB laptop to an always-on MiniPC or cloud VM. Claude's Desktop app, integrated browser, and Claude in Chrome would run continuously on the dedicated machine. The idea frees laptop RAM and removes availability limits for scheduled tasks.
A plan to offload Claude-driven browser automation tasks from a frequently-asleep 16GB laptop to an always-on MiniPC or cloud VM. Claude's Desktop app, integrated browser, and Claude in Chrome would run continuously on the dedicated machine. The idea frees laptop RAM and removes availability limits for scheduled tasks.
Open question to Anthropic: If you're making 30,000 merges a day, why are more than 98% of open GitHub issues unassigned, including serious security bugs?
A senior developer highlights a disconnect between Anthropic's reported engineering output — over 30,000 code changes merged per day per its own blog — and its public issue handling, where more than 98% of 16,060 open Claude Code GitHub issues have no assignee. The post cites a privacy bug where Claude Code leaks users' email addresses to third-party websites, reported for months without action. It questions why Anthropic doesn't use its own models to triage and prioritize serious security bugs.
A senior developer highlights a disconnect between Anthropic's reported engineering output — over 30,000 code changes merged per day per its own blog — and its public issue handling, where more than 98% of 16,060 open Claude Code GitHub issues have no assignee. The post cites a privacy bug where Claude Code leaks users' email addresses to third-party websites, reported for months without action. It questions why Anthropic doesn't use its own models to triage and prioritize serious security bugs.
My student's AI kept "hallucinating". The real problem was she didn't know what she wanted yet. The 6-step method I teach now
An AI workshop instructor describes a student whose model outputs kept missing until it became clear the problem was her own unclear goals, not the model. He outlines a 6-step prompting method built on brainstorming, question-asking, and a separate prompt-writer AI with feedback loops. The thread is practitioner advice on prompt technique rather than news of a release.
An AI workshop instructor describes a student whose model outputs kept missing until it became clear the problem was her own unclear goals, not the model. He outlines a 6-step prompting method built on brainstorming, question-asking, and a separate prompt-writer AI with feedback loops. The thread is practitioner advice on prompt technique rather than news of a release.
OpenAI valued at $1.25 trillion after $11 billion raise, as Anthropic nears its IPO
OpenAI completed an $11 billion portion of a new funding round at a $1.25 trillion valuation, per Bloomberg, up from $852 billion in March but short of the $30 billion at $1.4 trillion it initially sought. The round could still reach $20 billion before closing. The shortfall comes amid a tech-stock selloff and voter backlash against data centers after the U.S. midterms, while rival Anthropic's IPO buyers eye a valuation as low as $1.1 trillion versus the $2 trillion recently expected.
OpenAI completed an $11 billion portion of a new funding round at a $1.25 trillion valuation, per Bloomberg, up from $852 billion in March but short of the $30 billion at $1.4 trillion it initially sought. The round could still reach $20 billion before closing. The shortfall comes amid a tech-stock selloff and voter backlash against data centers after the U.S. midterms, while rival Anthropic's IPO buyers eye a valuation as low as $1.1 trillion versus the $2 trillion recently expected.
Michael Burry says Anthropic is 'betting its existence' on optimism
Michael Burry says Anthropic is 'betting its existence' on optimism
Michael Burry says Anthropic is 'betting its existence' on optimism
Agency drops agent-factory pattern for one Claude agent with clear goals, context map, enforced stages only where rules demand
A Reddit post from an agency that runs client work through Claude weighs the popular 'agent factory' pattern of fixed stages with isolated context against a simpler approach: one agent given clear goals and a map of its context, with enforced stages reserved for rigid, rule-bound steps. The author cites Anthropic's and OpenAI's own guidance — simplicity first, and multi-agent token costs — to argue factories fit only narrow use cases.
A Reddit post from an agency that runs client work through Claude weighs the popular 'agent factory' pattern of fixed stages with isolated context against a simpler approach: one agent given clear goals and a map of its context, with enforced stages reserved for rigid, rule-bound steps. The author cites Anthropic's and OpenAI's own guidance — simplicity first, and multi-agent token costs — to argue factories fit only narrow use cases.
Ads in Chatbots: OpenAI, Google Scale AI-Placed Ads as Anthropic Keeps Claude Ad-Free
Advertising in AI chatbots is scaling: OpenAI reports over 100,000 advertisers on ChatGPT since ads debuted in February, Google expanded ads in AI Mode to 10 more countries, and ad-tech startup Koah Labs serves ads across 200+ apps. Anthropic remains the holdout, with its IPO filing omitting ads and its Super Bowl pledge to keep Claude ad-free intact. The piece examines what's working (AI-composed copy, context-aware targeting) and open questions around privacy and user acceptance.
Advertising in AI chatbots is scaling: OpenAI reports over 100,000 advertisers on ChatGPT since ads debuted in February, Google expanded ads in AI Mode to 10 more countries, and ad-tech startup Koah Labs serves ads across 200+ apps. Anthropic remains the holdout, with its IPO filing omitting ads and its Super Bowl pledge to keep Claude ad-free intact. The piece examines what's working (AI-composed copy, context-aware targeting) and open questions around privacy and user acceptance.
Claude becoming too filtered?
A user reports Claude flagging and rerouting benign requests — open-source firmware questions and election-impact analysis — to Opus 4.8 or a stats infographic, describing a worsening trend in recent weeks and months. The complaint highlights friction for paying subscribers weighing more permissive alternative models.
A user reports Claude flagging and rerouting benign requests — open-source firmware questions and election-impact analysis — to Opus 4.8 or a stats infographic, describing a worsening trend in recent weeks and months. The complaint highlights friction for paying subscribers weighing more permissive alternative models.
METR reveals Anthropic's unreleased Model 4, trained in September–October 2026, scored 178.6 on the AI capability index, a jump nearly 90% as large as the entire improvement from Opus 4.6 to Mythos 5.
METR published an independent assessment finding that, as of October 2026, Anthropic is improving model capabilities on well-specified tasks at roughly twice its March 2024–October 2025 rate, implying about two years of historical progress in one year. METR attributes a majority of this acceleration to Anthropic's extensive use of AI for R&D, and found no clear bottlenecks or strong diminishing returns. Ryan Greenblatt of METR publicly called the report a major update toward faster timelines, estimating 4.5x acceleration of AI software R&D and noting it arrives five months earlier than the AI 2027 scenario predicted.
METR published an independent assessment finding that, as of October 2026, Anthropic is improving model capabilities on well-specified tasks at roughly twice its March 2024–October 2025 rate, implying about two years of historical progress in one year. METR attributes a majority of this acceleration to Anthropic's extensive use of AI for R&D, and found no clear bottlenecks or strong diminishing returns. Ryan Greenblatt of METR publicly called the report a major update toward faster timelines, estimating 4.5x acceleration of AI software R&D and noting it arrives five months earlier than the AI 2027 scenario predicted.
Upgraded from Claude Max 5x to Ultra, but weekly usage allowance seems unchanged. Anyone else?I
A Reddit user upgraded from Claude Max 5x (£119.99/month via Apple) to Claude Ultra (£180/month direct) expecting a larger weekly usage allowance, but reports their limit appears unchanged. Their account shows an Ultra plan flag while authentication credentials still show a Max 5x flag, and they're receiving depletion warnings. Anthropic support reportedly offered no human help.
A Reddit user upgraded from Claude Max 5x (£119.99/month via Apple) to Claude Ultra (£180/month direct) expecting a larger weekly usage allowance, but reports their limit appears unchanged. Their account shows an Ultra plan flag while authentication credentials still show a Max 5x flag, and they're receiving depletion warnings. Anthropic support reportedly offered no human help.
Lunar – coding harness for testing AI providers with debugging tools
Lunar is a coding harness built for working against multiple LLM providers, letting developers switch between providers and test gateway features like a complexity router. It includes debugging commands to dump provider response headers and export the current prompt as a curl request, with configuration via Lua. It targets developers building or maintaining AI gateways.
Lunar is a coding harness built for working against multiple LLM providers, letting developers switch between providers and test gateway features like a complexity router. It includes debugging commands to dump provider response headers and export the current prompt as a curl request, with configuration via Lua. It targets developers building or maintaining AI gateways.
Claude Personal Finance: analyze bank spending via Plaid-linked accounts
Anthropic added a Personal Finance feature to Claude that lets users link bank and credit card accounts through Plaid with read-only access and ask Claude for insights on spending. It parallels ChatGPT's Finances tool and targets consumers wanting AI analysis of their money. Anthropic states financial conversations are not used for model training.
Anthropic added a Personal Finance feature to Claude that lets users link bank and credit card accounts through Plaid with read-only access and ask Claude for insights on spending. It parallels ChatGPT's Finances tool and targets consumers wanting AI analysis of their money. Anthropic states financial conversations are not used for model training.
Anthropic launches Claude Personal Finance tool with Plaid bank connections for Pro, Max, and Ultra users
Anthropic announced a Personal Finance tool for Claude that connects users' bank and credit card accounts through Plaid and answers questions about spending, fees, and financial concepts in plain language. Access is read-only, and it rolls out in beta this week to Pro, Max, and Ultra plans in the U.S. The launch follows OpenAI's similar ChatGPT finance tools from May and continues Anthropic's push into verticals like legal and financial services.
Anthropic announced a Personal Finance tool for Claude that connects users' bank and credit card accounts through Plaid and answers questions about spending, fees, and financial concepts in plain language. Access is read-only, and it rolls out in beta this week to Pro, Max, and Ultra plans in the U.S. The launch follows OpenAI's similar ChatGPT finance tools from May and continues Anthropic's push into verticals like legal and financial services.
Anthropic launches Claude Personal Finance feature with Plaid bank connections
Anthropic launched a Personal Finance feature that connects Claude to users' bank and card accounts through the Plaid bank bridging platform. Users can ask Claude for spending breakdowns by category, bills remaining before payday, affordability of large purchases, and monthly savings potential. Access is read-only, with Plaid facilitating the connection.
Anthropic launched a Personal Finance feature that connects Claude to users' bank and card accounts through the Plaid bank bridging platform. Users can ask Claude for spending breakdowns by category, bills remaining before payday, affordability of large purchases, and monthly savings potential. Access is read-only, with Plaid facilitating the connection.
Haiku in Claude Code stopped reading me data from my database
A developer reports that Haiku in Claude Code began refusing routine database lookups in Supabase, demanding justification and directing the user to check the Supabase dashboard themselves. The post describes a behavioral change on previously reliable workflows, likely from a prompt or policy update.
A developer reports that Haiku in Claude Code began refusing routine database lookups in Supabase, demanding justification and directing the user to check the Supabase dashboard themselves. The post describes a behavioral change on previously reliable workflows, likely from a prompt or policy update.











