News for hackers. Stuff that matters.

stosssik

Breaking changes in third-party APIs are turning into a full-time job

For a few weeks I've been digging into breaking changes in the third-party APIs a company depends on.Nowadays, companies run more tools than they used to, and with time, it will drastically increase. First, because of vibe coding and apps built by agents, which can hallucinate an endpoint or a model name. And secondly, because there are more and more new tools in production for every team,...

( Read More... | 1 comment | on HN )
Re:SoftBank 2026 AGM [pdf] on Thursday June 25, 2026 @11:34AM (#48674992)
Original approach for their annual shareholder decks. I liek that.
Re: on Monday June 15, 2026 @09:22AM (#48540894)
Can't remember if the correct parameter for your LLM is max_tokens or max_completion_tokens ?

Say no more: we are launching modelparams . dev, the first complete model parameters database, open and collaborative.

You can browse it (UI), query it (API) or types in your app (NPM).

Re: on Saturday April 18, 2026 @09:23PM (#47821043)
A curated list of LLM APIs with permanent free tiers, no trial credits, no credit card traps.

Info included: rate limits, max context, and supported modalities.

Here's the list per provider:

Cohere (https://cohere.com/)

  • Command A (111B)
  • Command R+
  • Command R
  • + 3 more model (https://docs.cohere.com/docs/models)
Google Gemini (https://ai.google.dev/)

  • Gemini 2.5 Flash
  • Gemini 2.5 Flash-Lite
Mistral AI (https://mistral.ai/)

  • Mistral Small 4
  • Mistral Medium 3
  • Mistral Large 3
  • + 3 more model (https://docs.mistral.ai/getting-started/models/models_overview/)
Z AI (Zhipu AI) (https://z.ai/)

  • GLM-4.7-Flash
  • GLM-4.5-Flash
  • GLM-4.6V-Flash
Inference providers - Third-party platforms that host open-weight models from various sources.

Cerebras (https://cerebras.ai/)

  • llama3.1-8b
  • gpt-oss-120b
  • qwen-3-235b-a22b-instruct-2507
  • zai-glm-4.7
Cloudflare Workers AI (https://developers.cloudflare.com/workers-ai/)

  • @cf/meta/llama-3.3-70b-instruct-fp8-fast
  • @cf/meta/llama-3.1-8b-instruct-fp8-fast
  • @cf/meta/llama-3.2-11b-vision-instruct
  • + 5 more models (https://developers.cloudflare.com/workers-ai/models/)
GitHub Models (https://github.com/marketplace/models)

  • gpt-4.1
  • gpt-4.1-mini
  • gpt-4o
  • + 7 more models (https://github.com/marketplace/models)
Groq (https://groq.com/)

  • llama-3.3-70b-versatile
  • llama-3.1-8b-instant
  • llama-4-scout-17b-16e-instruct
  • + 7 more models (https://console.groq.com/docs/models)
Hugging Face (https://huggingface.co/)

  • Meta-Llama-3.1-8B-Instruct
  • Mistral-7B-Instruct-v0.3
  • Mixtral-8x7B-Instruct-v0.1
  • Phi-3.5-mini-instruct
  • Qwen2.5-7B-Instruct
Kilo Code (https://kilocode.ai/)

  • bytedance-seed/dola-seed-2.0-pro:free - Modality: Text | Rate Limit: ~200 req/hr
  • x-ai/grok-code-fast-1:optimized:free - Modality: Text (code) | Rate Limit: ~200 req/hr
  • nvidia/nemotron-3-super-120b-a12b:free
  • arcee-ai/trinity-large-thinking:free - Modality: Text (reasoning) | Rate Limit: ~200 req/hr
  • openrouter/free - Modality: Text | Rate Limit: ~200 req/hr
LLM7.io (https://llm7.io/)

  • deepseek-r1-0528 - Modality: Text (reasoning) | Rate Limit: 30 RPM (120 with token)
  • deepseek-v3-0324 - Modality: Text | Rate Limit: 30 RPM (120 with token)
  • gemini-2.5-flash-lite - Modality: Text + Vision | Rate Limit: 30 RPM (120 with token)
  • + 3 more model (https://llm7.io/)
NVIDIA NIM (https://build.nvidia.com/)

  • deepseek-ai/deepseek-r1
  • nvidia/llama-3.1-nemotron-ultra-253b-v1
  • nvidia/nemotron-3-super-120b-a12b
  • + 3 more models (https://build.nvidia.com/models)
Ollama Cloud (https://ollama.com/cloud)

  • llama3.1:cloud
  • deepseek-r1:cloud
  • qwen2.5:cloud
  • gemma2:cloud
  • mistral:cloud
OpenRouter (https://openrouter.ai/)

  • deepseek/deepseek-r1-0528:free
  • deepseek/deepseek-chat-v3-0324:free
  • qwen/qwen3.6-plus:free
  • + 9 more free models (https://openrouter.ai/models?q=free)
SiliconFlow (https://siliconflow.com/)

  • Qwen/Qwen3-8B
  • deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
  • deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
  • + 3 more model (https://siliconflow.com/models)
RPM = requests per minute • RPD = requests per day. TPM = Tokens per minute • TPD = Tokens per day • RPS = Requests per second • All endpoints are OpenAI SDK-compatible.

Free models you can use with your OpenClaw (no credit card needed)

We put together a list of free models you can connect to OpenClaw today through Manifest. No credit card, no trial that expires after 3 days. Just grab an API key and go.Here's what's available:- Google Gemini - 5 models including gemini-2.5-pro and gemini-2.5-flash. Up to 250K tokens per minute across all models. The pro model has a 1M context window on the free tier.- Cohere -...

( Read More... | 1 comment | 2 points | on HN )

Anthropic just cut off Claude subscriptions for OpenClaw

Just got the email. Starting tomorrow (April 4, 12pm PT), your Claude Pro/Team subscription no longer covers OpenClaw usage.` Hi,Starting April 4 at 12pm PT / 8pm BST, you’ll no longer be able to use your Claude subscription limits for third-party harnesses including OpenClaw. You can still use them with your Claude account, but they will require extra usage, a pay-as-you-go option billed...

( Read More... | 3 comments | 2 points | on HN )
Re:How to stop burning money on OpenClaw on Monday March 02, 2026 @02:19PM (#47222726)
I've talked to over a hundred OpenClaw users over the past two months. Cost comes up in almost every conversation. People set up their agent, use it normally, and end up with bills they didn't expect. $141 overnight from a misconfigured heartbeat. $800 in a month on a multi-agent setup. $30 burned doing barely anything.

The article digs into why this happens and what you can do about it. The core problem is that without optimization, every request hits your most expensive model, your system context loads on every call, and your conversation history grows with each exchange. It adds up fast.

The fixes range from simple config changes to architectural decisions. Routing tasks to the right model instead of sending everything to Opus. Using skills instead of spinning up multiple agents. Leveraging prompt caching on the provider side. Keeping your context lean. Running local models for lightweight tasks. And tracking costs daily instead of discovering a surprise bill at the end of the month.

Two deployments documented 77% and 80% cost reductions through these approaches. All sources and community reports are linked at the bottom. Happy to answer questions.

I mapped the major players building the OpenClaw ecosystem, just 2 months after its release.

What's unfolding around OpenClaw is unlike anything I've witnessed in open-source AI.

In 60 days: - 230K+ GitHub stars - 116K+ Discord members - ClawCon touring globally (SF, Berlin, Tokyo...) - A dedicated startup validation platform (TrustMRR) - And an entire ecosystem of companies, tools and integrations forming around a single open-source project.

Managed hosting, LLM routing, security layers, agent social networks, skill marketplaces.

New categories are taking shape in real time.

Some of these players are only weeks old. And established companies like OpenRouter, LiteLLM or VirusTotal are shipping native integrations.

Whether you're a VC exploring AI infra, an operator running agents, or a founder building in this space, this is the landscape right now.

Some of these startups are already pulling real revenue. Alternatives are stacking thousands of GitHub stars on their own. OpenRouter recently raised funding. The money and the users are already here. Most of what's on this map didn't exist 60 days ago.

This is what happens when an open-source project launches with the right building blocks at the right moment.

Ask HN: If your OpenClaw could do 1 thing it currently can't, what would it be?

Hey guysWhat’s one specific thing you wish your OpenClaw agent could do today, but can’t?Not vague stuff like “pay for things.” I mean which concrete use case ?For example:- “Automatically renew my AWS credits if usage drops below $100 and pay with a virtual card.”- “Find the cheapest nonstop flight to NYC next month, hold it, and ask me before paying.”

( Read More... | 3 comments | 4 points | on HN )
This resonates a lot. And we’re working on something in the same space: a way to build MCP aps for non technical people. If there are builders here who like experimenting, we’re looking for beta testers: -> https://manifest.build
Re:The future of software engineering is SRE on Monday January 26, 2026 @09:19AM (#46765921)
Totally agree. Vibe coding will generate lots of internal AI apps, but turning them into reliable, secure, governed services still requires real engineering, which is exactly why we’re building https://manifest.build. It lets non-technical teams build Agentic apps fast through an AI powered workflow builder while giving engineering and IT a single platform to add governance, security, data access, and keep everything production-ready at scale.

Older »