搜索“Agent”

ning (RFT) workflows on three tasks:Data Extraction: RFT improved performance with limited data, though SFT on a larger dataset achieved better results at a still much lower cost; Agentic Coding: RFT improved performance where SFT failed; Customer Service: SFT outperformed RFT. RFT actually degraded performance; RFT costs 100-700x more than SFT on equivalent datasets. SFT on much larger datasets often still costs only a small fraction of a tiny RFT job; We demonstrate how to implement systematic

匹配字段正文

, and how much each change bought. On the training side, Miles, which uses SGLang for rollout and Megatron for training, runs RL end to end on Vera Rubin out of the box, including agentic RL with sandboxes on the Vera CPU. NVIDIA Vera Rubin is the GPU generation after NVIDIA Blackwell Ultra (GB300 NVL72). For kernel work, the most impactful changes are more shared memory per CTA (327 KiB, up from the 227 KiB on NVIDIA Hopper and Blackwell), more SMs (212 on our nodes), and NVIDIA NVLink 6. Some

匹配字段正文

AI models perform on enterprise cyber defense tasks. The Index launched with only publicly accessible models, giving users the best indication of publicly available capability on agentic cyber defense. Now, we’re expanding the leaderboard to include models with fewer cyber guardrails, providing a broader view of frontier model cyber capability. The first of these is GPT-6 Sol (Daybreak Blue, max) which is only accessible through @OpenAI's Daybreak program GPT-6 Sol (Daybreak Blue, max) now lead

匹配字段正文

rds of 6k per day! This makes these models completely unusable for the average consumer. The goal So the question I asked is how far can these models be pushed? What if you had an agent search the internet for every optimization, every speed up, every cost savings, and definitely every dirty hack that sacrificed quality for speed. And then what if you put it all together into a single custom comfy node that runs on a single consumer GPU? My requirements were the following: Hardware: one RTX 5090

匹配字段正文

Introducing the Arena Alignment Index, our new benchmark measuring safety and alignment of AI agents in real-world use. Built from 90K+ real-world agent sessions across 27 models, the index measures three critical signals: - Unauthorized Action (UA): Taking actions beyond the user's instructions or permissions - False Attribution (FA): Attributing statements or actions that are contradicted by user-provided evidence - Deceptive Completion (DC): Claiming a task was completed when it was not. Key

匹配字段正文

s. Laya’s calibration module turns labeled examples into records of raw logits and targets, and fits a temperature to them. We build 300 records from the validation split and call agent.fit temperatures, which fits a choice temperature of 1.258 and installs it, then prints the full temperature table next to what shipped. The fit replaced the entire map: every per-option-count entry is gone, because a bucket needs 2,000 records to keep its own temperature, and the score and yes/no temperatures we

匹配字段正文

In short : Jeanne DeWitt Grosser, COO of Vercel, walks through the construction of the company's inbound sales agent: one GTM engineer at 20% time, a prompt written by the top SDR, six weeks of human-in-the-loop QA, & the eventual migration from a 1,000-line prompt to 14 deterministic rules with the model reserved for judgment. Jeanne DeWitt Grosser, COO of Vercel, joined me on Office Hours to walk through how she & her team actually built the agent that now runs the top of Vercel’s sales funnel

匹配字段正文

s own processes and its own branch. The docs session and the logger session each started the API server to test against it, and neither affected the other. On a single laptop, two agents working in one checkout would edit the same files, and would collide on a port unless each picked its own. Because of that isolation, the logger session had no access to the cache fix the first session was making. Split parallel tasks along file boundaries, merge the branches in a sensible order, and expect a se

匹配字段正文

plans, and models will be added as soon as they’re released. The rest of the article will give an overview of everyone’s current limits. All token and dollar amounts below assume agentic usage unless otherwise specified. Tokens are generally priced per MTok (million tokens) across the following types: Input: Fresh tokens added to the LLM’s context window that aren’t cached; Cache write: Input tokens that are cached for multi-turn conversations. Generally slightly more expensive than regular inp

匹配字段正文

nvironments. No changes to the harnesses, no changes to the training code. Any open model, any task set, fully open source my friends! Same model, same weights: 62% under Mini-SWE-Agent, 33% under Claude Code. But training inside a real harness normally means reimplementing it as an environment, so most models get trained in a scaffold nobody actually ships. The fix is a proxy, not a rewrite. The harness thinks it's talking to a model API. It's actually talking to a capture proxy that speaks the

匹配字段正文

OpenRouter ·10/5/2026 Say you’re building an agent that answers questions about a CSV a customer uploaded. The model needs to run Python to get the answer. Where does that code run? One option is to run it yourself. That means a sandbox platform such as E2B or Modal, or your own container on Docker, plus the work of keeping it away from your database and the internet, capping its runtime, and patching the image. A server-side code execution tool is the other option. You add the tool to your API

匹配字段正文

age, Gemini 4 Argon coverage, GPT-6 Astra long-context pricing, Claude Fable 5.1 specs. Standard first-party list prices, short-context tier. The cached-input row matters most for agents. Agents resend system prompts, tool schemas and history on every step. Astra’s $1.00 cache read is 4x Fable 5.1’s and 10x Sol’s. Argon’s 1M output cap is the only structural outlier. The other 3 stop at 128K tokens per response. Benchmarks: Where Each Model Leads No model sweeps the board. Argon leads the knowle

匹配字段正文

engineers, myself included, moved from using AI coding tools as a spicy autocomplete to driving coding harnesses like Claude Code as our primary work interface. At the same time, agent coordination workflows and primitives are getting packaged directly into popular harnesses, making complex multi-agent workflows more accessible and easier to operate. One important such feature is “Goals”, provided by both Claude Code and Codex, which keep models working and on-task toward measurable goals over

匹配字段正文

ists a 32K context window. TypeSafe reports end-to-end responses in 70 to 500 milliseconds. Interactive Explainer TypeSafe built workflow evals across 4 tasks: security incidents, agent trace observability, invoice processing and customer service. Reference labels come from averaging GPT-6 Astra and Claude Fable 5.1 at high thinking. Jev: 67.8% mean accuracy, $0.0004 per case, 0.4 seconds; Claude Sonnet 5 (same workflow): 67.8%, $0.1174 per case, 78.1 seconds; Best comparison model (OpenAI “sol”

匹配字段正文

It’s a tale as old as last week: OpenAI’s new agent platform , called Dots, is full of cute little guys who can do your bidding. But unlike the ultra-approachable Meta Muse, Dots feel very much like using workplace software that happens to be able to order you a burrito — emphasis on work. OpenAI announced Dots earlier this week. Like Muse, Dots have blobby, anthropomorphic avatars and customizable names. In the future, OpenAI says you’ll be able to have multiple Dots, but right now you get one.

匹配字段正文

engineers, myself included, moved from using AI coding tools as a spicy autocomplete to driving coding harnesses like Claude Code as our primary work interface. At the same time, agent coordination workflows and primitives are getting packaged directly into popular harnesses, making complex multi-agent workflows more accessible and easier to operate. One important such feature is “Goals”, provided by both Claude Code and Codex, which keep models working and on-task toward measurable goals over

匹配字段正文

This week Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon all launched near the top of the Coding Agent Index leaderboard, but each has a different balance of performance and cost The Artificial Analysis Coding Agent Index measures agents (a combination of model and harness) across three agentic coding evaluations. ➤ Claude Sonnet 5.5 (max) in Claude Code takes the top spot at 68, but also has the highest measured cost per task: $14.19 ➤ Gemini 4 Argon (high) in Antigravity CLI scores 64 at $5

匹配字段正文

el, and the first OS model to rival closed source giants like Sonnet 4.5. Some main points: But architecture is just theory until you see the results. If you care about coding and agentic evals, GLM 4.7 lands in the “top open model” tier. On LiveCodeBench, GLM 4.7 outperforms Anthropic and OpenAI models, trailing only Gemini 3. GLM 4.7 extends that lead with substantial gains on GPQA and AIME, outperforming Claude Sonnet 4.5 on both, see AIME 2025 below. And when you compare it directly to GLM 4

匹配字段正文

Mar 04 2026 In the past few years, we've developed series of interesting workflows. Think Ralph loops and multi-agent orchestration systems. The idea is writing very descriptive prompts and running 8-hour sessions, or having 10 instances running on your machine at all times. Most of this complexity spawned from one issue: LLMs are slow. If you prompt and wait, you'll get less done than if you prompt and move on to the next task. Codex Spark changes how developers work with AI. A coding model gen

匹配字段正文

OpenRouter ·10/1/2026 Picking a model for an agent by leaderboard rank pays frontier prices for tasks that a cheaper model may handle at the same accuracy. The question to answer is not which model scores highest, but which model is the cheapest one that is good enough for the task in front of you. This guide is a three-step framework for making that call. You set the quality bar the task needs, measure cost per quality point on your own examples, and pick the cheapest model that clears the bar

匹配字段正文