model family hub

Claude models

Anthropic model family used by Claude products and the Anthropic API.

Published changes239StatusactiveAliases3Relations1

Names and model IDs

ClaudeClaude APIclaude-api

Source-linked relationships

Timeline

239 published changes

Open filtered feed
Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Community Importance 32/100 Confidence 55/100

Show HN: Klistra – Make launch videos and product demos, all in one AI studio

I built Klistra: describe your video idea and add your logo, photos or other assets, Klistra designs and animates a video for you, exported for every screen size, with fonts, media and logos kept exact. Deterministic renders make features exact and isolated, allowing for superior control and iteration compared to diffusion models. Here's a short demo render where I made an ad for Klistra itself: https://app.klistra.ai/s/4EQoWCe37SkYz5nThgegzw As a solo dev with a lot of projects going on, I found myself needing to make short promos or explainer videos. Not being familiar with animation software, I had the idea to use Claude to code up those videos and export them as video files. In the process of creating these clips I stumbled into a bunch of bugs and setup needed to make it work reliably. The list of features I wanted for better control and feedback kept growing and I needed to sign up for a bunch of services to get music, voiceovers and other media into my project. I couldn't find an existing service that had everything I wanted. Seeing these hurdles and that the technology was only easily accessible to people already confident with coding agents, I started working on Klistra, a

Release Importance 76/100 Confidence 95/100

v2.1.330

## What's changed - Added sign-in to organizations managed in Anthropic's Enterprise Admin Console: after `claude auth login`, Claude Code applies the organization's settings and uses its cloud provider or gateway - Added easel, a built-in mod that draws charts when asked and under MCP metrics results, on Pro, Max and, where available, Team and Enterprise on macOS and Linux; turn automatic charts off in `/config` - Added `CLAUDE_CODE_FULL_SIGN_IN_LINK=1` to print the full sign-in link, if the short one shows an error or a script reads values from it - Self-hosted runner: Added `--start-dir` / `SELF_HOSTED_RUNNER_START_DIR` to have Claude start its work in a folder of a repository, while settings and hooks load as without it - Added a notice in cloud and self-hosted runner sessions when a plugin marketplace can't be fetched because its repository isn't one of the session's - Self-hosted runner: Added a notice when Anthropic-managed git refuses a plugin's own repository, naming the plugin and what to do - Added `commandPrinted` to `claude plugin install --json` results, true when the plugin's install command wrote to stderr, and `commandFailed` when it exited with an error, was kille

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Integration Importance 64/100 Confidence 95/100

v2.1.329

## What's changed - Added a trust dialog warning that says where a folder's hooks are declared - Added a warning in the conversation, in Claude Desktop and other non-interactive hosts, when hooks in managed settings fail to load - Added this session's scheduled tasks, `/loop` wakeups, the goal and this conversation's forks to `/tasks`; x stops one after asking - Added a "Server message" row to a server's page in `/mcp` that shows the JSON-RPC error message of a remote HTTP MCP server that refuses the connection - Added a "Marketplace auto-update" row to `/status` that names managed marketplaces that get no automatic updates, and why - Added support for turning on Code Mode while OpenTelemetry events and metrics are enabled; each tool call the code makes sends its events; tracing still keeps it off - Fixed `~/.claude.json` sometimes being reset to defaults when a network stall or file lock made it unreadable and cleared during a save - Fixed repeated sign-outs on macOS when Claude Code could read the Keychain but not save to it. That session now stops at expiry and says why, and other sessions stay signed in - Fixed `/terminal-setup` corrupting VS Code settings or keybindings files

Community Importance 32/100 Confidence 55/100

Show HN: Zpup – Zdog style 3D scenes in JSON, SVG from any angle

Hi. I made small project called Zpup. It takes small 3D scene written in JSON and gives back a flat SVG from any angle. There is no JavaScript in the file, so it works in img-tag or README. I needed 3D for small project at my company, and Zdog make big impression on me. But Zdog has not had a release for several years, so I wanted to make modern version. One thing works different. Zdog sorts each shape by one point. so small thing on big floor gets covered when you turn it. Zpup compares the shapes that overlap. You can see both in the README. I wrote it with Claude Code Fable 5.2. No canvas renderer, no text shape, and it is 0.2 version. Playground: https://rxolve.github.io/zpup Feedback is always welcome :)

Community Importance 32/100 Confidence 55/100

Show HN: Git Hook for Model Identifation

I have started using multiple models (Claude, Codex, Grok via cursor-agent, Z.ai, Deepseek, Hy4 via OpenCode, agy) etc, and I realized that I not able to identify which commit was made by which model. After looking around and not finding anything, I create a format specification that will let Model self-identify itself. The goal is a higher ranked model can verify the output of other models and suggest changes. It is similar how to a Senior Engineer will review the output of Junior Engineers. As Humans, we know which model is behaving well, and which we can trust more. My AI based commits have a identification like this. Add retry backoff to fetch Model: claude-sonnet-5-5 harness=claude-code effort=high conf=medium Model: opencode/hy4-flash-free harness=opencode effort=high conf=medium In using this, I get an intuition what I can expect when I fire off model on a particular task.

API change Importance 78/100 Confidence 95/100

v2.1.328

## What's changed - Added auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and Claude apps gateway; interactive sessions with no permission mode start in it - Added `"onFailure": "block"` for `mcp_tool` hooks, as for command and HTTP hooks: a hook whose call fails blocks the action - Added a warning when a message from a session that skips permission prompts may be held and dropped by a cloud session - Added a notice in `/status` and `claude doctor` when an OpenTelemetry exporter has no endpoint and its data would likely be lost - Added `CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME`, which lets a host stop `claude -p --resume` and `--continue` from running a tool call that a hook deferred - Added support for apps that embed Claude Code to switch on the policy plugin's `allowManagedModsOnly` option; an admin's managed settings still win - Added a startup warning to the Claude apps gateway when a policy's `maxEffortLevel` is not a valid effort level - Self-hosted runner: Added `--repository-folder` to name the folder that sessions use for a repository at `--capacity 1`, in place of its path under the base folder, which can't

Community Importance 32/100 Confidence 55/100

Show HN: Robotics needs better CAD, so I wrote a B-rep kernel and 3,487 tests

Hey HN, I’m Tony. I’m a co-founder of Springcraft, where we have been doing a lot of CAD modeling for robot parts. Our team kept running into limitations with existing kernels, and so I decided to write a new B-rep kernel and a new comprehensive, kernel-agnostic B-rep conformance test suite. Both are available today under MIT licenses. The kernel, springrep, is written in safe Rust with no dependencies. It covers a broad range of our mechanical CAD needs, including advanced fillets, sweeps, lofts, shells, offsets, STEP and exact measurements and queries. Claude Code made it feasible to write this, but it took a lot of time and care to get it right. The test suite, brep-conformance, has 3,487 cases covering the modeling operations common across kernels: booleans, fillets, chamfers, shells, offsets, sweeps, lofts, STEP, and queries with expected answers coming from closed forms and exact geometry. It caught hundreds of bugs, edge cases, and hot spots in springrep that our own tests had missed, and now springrep passes all of it. We’ve also run it on a few popular open-source kernels and anyone can run it on theirs. This is the first version, so there are surely mistakes in the test a

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Community Importance 32/100 Confidence 55/100

Show HN: Terminess, a harness that runs the official Claude Code and Codex CLIs

TLDR: I built a terminal (harness) for Claude Code and Codex that works the official way, on your own subs. Anthropic stopped Claude subs working in 3P harnesses in Feb, 9 months of suffering later, I decided to build one that plays by their rules, a terminal (harness) that runs Claude Code and Codex (and others), together. (I am not claiming that this is a Deepseek harness killer lol, but it works really well for me) I use a number of agents in tandem, on my laptop and my server. Checking them from my phone took 4+ apps and I was losing hours away from my laptop, and copying context between them ate the rest. Terminess puts it all in one place. It complies with all their terms. Terminess runs the real claude and codex CLIs, unmodified, signed in with your own plans, exactly as your reg mac terminal or iTerm would. It doesn’t read, store or reuse your login tokens, and never proxies your requests, so nothing about it works around Anthropic's or OpenAI's terms (currently). Underneath is its own terminal engine, written from scratch for agent work. - Sessions live on your Mac or your own server, so agents keep going with the lid shut - You can start on your phone and pick it up on yo

Pricing Importance 82/100 Confidence 95/100

v2.1.327

## What's changed - Added a confirmation before `/install-github-app` changes a repository - Added `claude mcp add-json -` to read the server's JSON from stdin, so tokens can stay off the command line; not available inside a Claude Code session - Added a line when a resumed session's earlier messages can't be loaded - Added a list in `/status` of broken agent, skill, command and output style files, keybinding errors, an unavailable sandbox, ignored permission settings and hook matchers that never match - Added a note in `/config` when other settings override a row - Added `x-claude-code-managed-settings-checksum` to the gateway hint headers; sent where managed settings set `CLAUDE_CODE_GATEWAY_HINT_HEADERS=1`, including where they already do - Added support for naming a session's billing group with the `attributionProject` setting, for Claude apps gateways that set `admin.billing_group_tag` - Added `/tui classic` as the name for the non-fullscreen renderer; `/tui default` still works - Remote Control: Fixed stopping the server in worktree mode deleting a running session's git-ignored files, such as `.env` - Fixed self-hosted runners with `--use-anthropic-git-proxy` deleting a non-e

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Integration Importance 64/100 Confidence 95/100

v2.1.326

## What's changed - Added support for turning on Code Mode while agent teams are enabled; if you already turned it on, sessions with teams on now start in Code Mode - Added a startup warning when an `--allowedTools`, `--disallowedTools` or `--tools` value has an entry that doesn't look like a complete rule - Added a message explaining why MCP servers stop loading when an administrator adds `managed-mcp.json` during a session - Added a row to `/status` that says when your provider or an environment variable such as `DISABLE_TELEMETRY` turns features off - Added a note in `/plugin` and `claude plugin list` when a malformed `plugin.json` field such as `author` or `homepage` is ignored - Added Ctrl+O to show hidden startup notices - Added a hint in the empty prompt that says what to do after a request fails - Added an `esc to stop` footer hint in a subagent's view where Esc would stop it - Added `OTEL_RESOURCE_INCLUDE_HOST` to include `host.name` and `host.id` in OpenTelemetry resource attributes (off by default) - Added `overflowX="scroll"` on a mod's `Box`, and `maxLines` and `wrap="clip"` on its `Text` - Fixed `cp` of a file being refused when only its source was covered by an `Edit

Community Importance 32/100 Confidence 55/100

Show HN: Fakemetrics.com – "build in public" like a superstar

Hi HN, I am Tigran (github.com/tigranbs) I made Fakemetrics.com as a fun experiment while testing how well Claude Opus can build pixel-identical UIs. It turned out it wasn't as easy as just prompting it, but I spent time building the system so it works almost every time. And I am pretty sure you are seeing a bunch of people posting weekend vibe-coded apps and saying "$50K/month after quitting my McDonnalds job...". Well, even my friend tried to fake it a few months back. So, I kinda did this fakemetrics.com as an experiment to see how well I can have the system that clones UIs, and packaged that as a very simple (I think) usable product for the people that want to be a superstar "build in public" type of a person with no revenue, but a perfect Stripe user growth screenshot. Useful? - Probably not that much Fun? - YES!! The best part is that Claude, ChatGPT, or Grok models aren't that good at UI clones by default, but with fakemetrics.com I kinda made a system that produces near-identical results from screenshots fairly fast. That was worth learning! Let me know what you think? Should I move on with the UI cloning as an open-source product?

Community Importance 32/100 Confidence 55/100

Does Anyone Use Happier.dev?

It seems to provide a way to have a single view of sessions across multiple agents and harnesses - I'm planning to use it so I can use Claude Code and Goose + OpenRouter in one place, and possibly orchestrate activity across the two. Just wondering whether anyone has any experience of it, what your thoughts are, and whether there are alternatives I should be looking at?

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Community Importance 32/100 Confidence 55/100

Ask HN: Is post-training make AI models dumber?

Post-training is making models better at benchmarks. They seem more consistent at coding, but dumber to me and I want to know what y'all are noticing. Here are the big three I'm noticing (mostly in claude): 1) Agents cannot write natural language well. Claude uses vague phrases to avoid any technical term such as "What the caller calls" instead of saying "callee". 2) Agents cannot manage encapsulation; they always make spaghetti code. 3) Agents are afraid to break anything. They always add more code to work around the code that's already there rather than restructuring it. No amount of rules or coaching helps much. Ask an agent to make code simpler and it usually gets longer. Add any sort of rule, and the agent will follow it like you just got a wish from the monkey's paw. Has it always been this way or is post-training compressing the output to predictable failure modes?

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Community Importance 32/100 Confidence 55/100

Show HN: Feedback/AMA MeditionBench; LLMs for mediation and conflict resolution

I am the founder of Human Assisted Intelligence and I believe AI can be used to reduce polarization and improve the way people communicate and reach agreement. At https://mediationbench.com/ , there are many frontier LLMs tested, but it is difficult to keep up. I've been trying to get more free openrouter or huggingface credits. Azure for OpenAI and Anthropic have both provided token credits. (Thank You!) The research paper is there as well. There have been efforts to use LLMs to lie and play diplomacy, which are excellent compliments to this work. In v4, the test will get harder as Claude nearly saturated the test. Astra was surprisingly low, perhaps because of alignment or safety efforts. I'd love to get feedback and hear from all of you. I've been working on related ideas for 20 years, so it's definitely a passion project with too much time in isolation.

Tutorial Importance 42/100 Confidence 95/100

v2.1.325

## What's changed - Fixed a security issue in how zstd-compressed responses are decoded - Added a spinner tip in VS Code and similar integrated terminals about re-running a notebook cell, shown once Claude has read or edited a Jupyter notebook - Added a line under the error shown when your organization's policy blocks a request, saying how to continue: run `/rewind` and change the message, or run `/clear` - Added a note for Claude when a sandboxed `git checkout`, `git stash` or autostash pull cannot update protected files such as those in `.claude` - Fixed `/security-review` and skills' inline git commands stopping with an error in worktrees of bare repositories, `--separate-git-dir` checkouts and Android `repo` checkouts - Fixed subagents with their own model being moved to the plan-mode model when the model setting is `opusplan` or `haiku` - Fixed subagents on a different model sizing file reads and MCP results with the main session's model - Fixed sessions on many third-party gateways and Foundry stopping with a context length error on every turn instead of compacting - Fixed recent turns sometimes going missing from a resumed session after Claude Code stopped without a clean ex

Community Importance 32/100 Confidence 55/100

Show HN: Langfuse AI Gateway, open source rust gateway to trace every LLM call

Hi HN, this is Max, Marc and Clemens, co-founders of Langfuse, the open source tool for agent & observability & evals ( https://github.com/langfuse/langfuse ). It’s been a minute since our last Show HN in August 2023 ( https://news.ycombinator.com/item?id=37310070 ) - we’ve been busy building out and serving our tracing & eval pieces over the past years and have been very lucky to build a successful company and open source project around it. Today, we are happy to show the beta of Langfuse AI Gateway: https://github.com/langfuse/langfuse/tree/main/ai-gateway . You may ask, why an AI Gateway as our second product: it’s been one our community’s strongest asks + integration pattern that we’ve seen. We want to offer a scalable, reliable and secure way to solve for the gateway use case in the same way that we’ve done with tracing/evals over the past years. And we’ve recently grown our team with some excellent folks - expect to see more bigger bets from us. Our observability SDKs currently see the code you instrument. Much of a company's model traffic runs through code that is not or cannot be easily instrumented: coding agents like Pi, Claude Code and Codex, internal tools, scripts etc.

Community Importance 32/100 Confidence 55/100

Show HN: Carton Clock – A working pendulum clock made from empty drink cartons

After my last trash clock ( https://news.ycombinator.com/item?id=49930439 ), I wanted to see how low I could go. My latest clock is built mainly from empty drink cartons, plus a few skewers, some paper straws, a piece of string, some nuts as weights, and lots of hot glue. It has a pendulum, an escapement, and hands for seconds, minutes, and hours, and runs for about an hour. The design of the gears and escapement was AI-assisted: Claude built me interactive browser tools that generate templates and 3D-printable jigs and tools. I’m happy to answer your questions! And to the commenter who suggested I submit trash clock timekeeping as a talk to 40C3: I did, and it got accepted. CU at Congress, I guess!?

Community Importance 32/100 Confidence 55/100

Show HN: Avatar/pet plugin changes attire based on topic Codex/Claude/DSH

Avatar/pet plugin for Codex/Claude/DSH/Doubao/Yuanbao... It will change the attire to look like an expert the topic/question you asked. it runs fully offline, using a tiny local model (embeddingGemma). Free and open source. Love to hear your comments. https://github.com/hci-yeo/AdaptiveAgentAvatar

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Community Importance 32/100 Confidence 55/100

Ask HN: What models do you use for tasks Anthropic/OpenAI don't approve of?

People frequently talk about the ever growing lists of tasks that claude and friends will "refuse" to work on, and then frequently someone will make a comment about "open models" or even specifically uncensored/"abliterated" or something, but are rarely specific. If you just search you can find a lot of models claiming some level of "uncensored" but presumably you also need them to be some level of capable. So, for those of you who need to do "unapproved" things, what are your go-tos?

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Integration Importance 64/100 Confidence 95/100

v2.1.321

## What's changed - Added Claude Fable 5.2 (`claude-fable-5-2`), now the default Fable model — 1M context, $8/$40 per Mtok with $0.20/Mtok cache reads - Added `--settings ` to `claude remote-control`, which applies a settings file to every session it starts; if the file changes, no new session starts until you restart `claude remote-control` - Fixed MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them - Fixed a case where a read-only Bash command ran outside the sandbox without a permission prompt - Fixed a case where Bash permission checks judged a file write from the wrong folder and didn't prompt - Fixed deny rules and protected-file prompts not applying to the file that a `cp` or `mv` into a directory writes - Fixed the "Draws from usage credits" note showing after a switch to Fable on Enterprise seats whose plan includes Fable - Fixed the fallback model sometimes declining a request or mentioning a stopped response when the first model was blocked before replying - Fixed `claude mcp login` printing a SEP-2352 issuer warning each time you sign in again to a server that Claude Code registered with - Fixed `NotebookEdit` deleting

Community Importance 32/100 Confidence 55/100

Show HN: Native Tcl/Tk Wayland and Android Port

Native Wayland for Tk was proposed years ago as a hard GSoC project ( https://wiki.tcl-lang.org/page/GSoC+Idea%3A+Tk+Backend+for+t... ) - and let's be honest, it's very hard for humans. Turns out AI cracks it in a few hours. Claude was given the freedom to make its choices from top to bottom, including stack and architecture. It came up with a thin X11 emulation layer, the same approach Tk's macOS and Windows ports already take - except this one is shared by Wayland and Android thanks to their similarity. The human in the loop merely tested and sanity-checked the outcome to her unexpected satisfaction. Tkabber (a large, long-lived XMPP client) runs on the Wayland port unaltered, and on Android with one small fix. Unlike AndroWish (an SDL-based Tcl/Tk distribution for Android), this is a backend inside Tk itself, shared with the Wayland port: Cairo/HarfBuzz drawing, with touch and the on-screen keyboard handled in generic Tk through new commands and events. If you want to experience an anachronism, you can try the APK of the Tk demo with an added multitouch example ( https://github.com/pounceandmiss/tk-wayland/releases/ ), and full-on Tkabber on Android ( https://github.com/pouncean

Community Importance 32/100 Confidence 55/100

Show HN: NLP capabilities of Opus, Astra, Argon measured by Codenames skills

Hello Show HN, I'm a college student, and I enjoy playing word games like Codenames, but it can be hard sometimes to find enough people to play when everyone is available, so I made an app that allows you to fill any seat with an AI teammate and also play against AI. This way it allows for people to get the full strategic experience of codenames just on their own. Through this experiment, I was also able to measure and benchmark a few different LLMs through AWS Bedrock API calls and give them ELO ratings based on stuff like win rate and clue/guess accuracy. If anyone is interested, they can see more about my benchmarking methodology here: https://wordspies.com/hn#method . I think it's pretty interesting because it gives you a lot of insight into these various LLMs' NLP capabilities. A few interesting stats were that I rated Gemini 4 Argon at 1775, GPT-6.1 Sol at 1744, GPT-6 Astra at 1723, Claude Fable 5.1 at 1623, and Claude Opus 5.5 at 1592. You'll find a few interesting things there, like how, although newer, more expensive models tend to perform better, some rare exceptions exist, like Kimi K2.5 slightly edging out Kimi K3. Check out the site if you're interested in NLP and data

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

Community Importance 32/100 Confidence 55/100

Show HN: Conductor sold to Meta. Agents rebuilt the app in 3 days for $170 (MIT)

Re: https://news.ycombinator.com/item?id=50369503#50370152 --- There are still a few kinks to be worked out, but Conductor 0.92.1’s own frontend can run on this project’s Rust shell and sidecar. Every one of its on-screen strings, native commands and sidecar methods is matched or listed as a deliberate difference. On scripted scenarios, this project’s sidecar produces the same events as Conductor’s. Deliberate differences are documented in docs/decisions.md. Built with Opus 5.5 orchestrating Deepseek-flash agents with Umbel and Pleach (see plotplot.ai) with Sonnet reviewing each batch and Opus fixing what the reviews found. Size: about 197,000 lines of TypeScript and Rust, not counting tests. Cost: About ~50 USD of DeepSeek plus ~70% of a weekly Claude 20x subscription. PS: Oh, and it duplicates some of their cloud functionality by running agents on cloudflare workers/containers

Community Importance 32/100 Confidence 55/100

Show HN: OS Grokbot, can run on Claude/codex subs

Made a provider-agnostic, entirely locally hostable GrokBot that wraps around your tmux-based terminals - which means, your claude/codex/etc. subscriptions can run inside of it. It is better at long-term tasks, business workflows, and operational work. We have not yet had a chance to run the benchmarks.

Product update Importance 82/100 Confidence 100/100

DeepSeek-V4.1-Flash-1017 Update ​

DeepSeek-V4.1-Flash-1017 Update ​ DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model. Performance–cost Pareto improvement Compared with the previous DeepSeek V4.1, the new model delivers substantial improvements across benchmarks: HLE (w/o tools): 42.6 (44.0*) Terminal-Bench 4.0: 42.3 DeepSWE v1.1: 77.0 ProgramBench: 29.1 NL2Repo-Bench: 68.2 Automation-Bench v1.0.6: 60.5 Agents' Last Exam: 35.4 Apex Bench v1.1: 72.1 Chartography (w/tools): 83.0 BabyVision (w/tools): 92.5 ZeroBench-main (w/tools): 51.0 * Tested only on the pure-text subset of the HLE benchmark set. Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5. While improving performance, we also reduced the model's task runtime and output length, enabling users to complete tasks at lower cost and higher speed and achieving a new performance–cost Pareto optimum across multiple harnesses. We specifically optimized for DeepSeek Harness, Ten

API change Importance 78/100 Confidence 95/100

v2.1.319

## What's changed - Changed auto mode with telemetry or feature flags off, including behind LLM gateways: calls the server can't check are asked about or denied, with no fallback to the built-in classifier; set `CLAUDE_CODE_AUTO_MODE_SERVER=0` to keep using it (Bedrock, Vertex and Foundry are unchanged) - Changed auto mode behind an LLM gateway: `CLAUDE_CODE_AUTO_MODE_SERVER=1` turns on the server-side check even with `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS` set; if managed settings set the latter, they must set the `1` too - Added support for MCP tasks in the terminal: when an MCP server runs a tool call as a task, Claude Code moves it to the background right away and tells Claude the result when it finishes - Added the `CLAUDE_CODE_DISABLE_MCP_TASK_BACKGROUND` environment variable to make Claude Code wait for an MCP task inside its tool call instead of moving it to the background right away - Added `permissions.denyInsteadOfPrompting` setting: a tool call that would wait on a permission prompt is denied and the turn continues, in terminal sessions, subagents and teammates - Added `/claude-api latency-optimize` to find where your Claude API code makes people wait and try faster se

Action requiredReview model IDs, endpoints, and client configuration before the change reaches production.