# OpenAI Codex Harness as an API > Source: > Published: 2026-09-11 12:30:00+00:00 - [unwind ai](../) - Posts - OpenAI Codex Harness as an API # OpenAI Codex Harness as an API ## + Cursor Projects to direct 1000s of agents ### **Start here ↓** ### OpenAI turns the Codex harness into an API Until now, building a serious agent meant assembling the difficult parts yourself: a harness, sandboxes, context management, tool routing, and a way to coordinate subagents. OpenAI just bundled that machinery into the Agents API, the same managed harness and infrastructure behind Codex. A single API call can now launch an agent with its own compute environment, files, tools, MCP servers, and up to three concurrent subagents. It can keep working across multiple context windows, automatically compact old context, find tools only when needed, and combine tool results in code instead of dumping everything back into the model. The API is in public beta with no platform fee; you just pay for models and tools. ## 🚀 **Shipped** **Sakana AI says a team of cheaper models can outrun the frontier.** They launched Fugu Max and Fugu Ultra v2, which route work across a swappable pool of open and specialized models instead of throwing the biggest model at every problem. Fugu Max approaches elite-model performance at 2-6 x lower cost, while Ultra v2 beats Opus 5 and Fable 5 on its Chartography benchmark, without using Fable 5, Fable 5.1, or GPT-6 Astra.[Sakana’s announcement](https://sakana.ai/fugu-max-release/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api)  **A 35B model now runs on an iPhone by treating its SSD like memory.** Edge0 runs sparse mixture-of-experts models on Apple Silicon by streaming weights from the SSD instead of loading the entire model into RAM. Its 35B model occupies 23 GB on disk but uses around 2.9 GB of active memory, reaching 14.9 - 17.7 tokens per second on an M4 Pro Mac mini. Squeezing this much model into this little live memory is wild!![Edge0 on GitHub](https://github.com/Edge0-AI/edge0/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api)  **DeepSeek introduced V4.1-Flash, the smallest model in its new architecture family,** with native visual understanding and a 552B-parameter MoE design. It is smarter, faster, and cheaper than the previous generation. Its dramatically smaller cache uses one-quarter the HBM and one-eighth the SSD storage. Live through the DeepSeek API as `deepseek-flash` with lower peak pricing and another 50% discount during off-peak hours.[DeepSeek announcement](https://x.com/deepseek_ai/status/2097930608790167907?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Google bundled their entire Cloud expertise, live docs, and guardrails into one plugin.** This Cloud Developer plugin gives coding agents the skills and tools needed to handle authentication, permissions, projects, and `gcloud` operations safely. Because it follows an open Agent Plugins specification, the same bundle can work across Antigravity, Claude Code, and Codex.[Google Cloud’s announcement](https://cloud.google.com/blog/topics/developers-practitioners/introducing-the-google-cloud-developer-plugin-for-ai-coding-agents?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Cursor launched Projects, a persistent coordinator that can remember months of context**, delegate to thousands of subagents, and react to Slack messages, schedules, or pull requests without waiting for a prompt. Their new users merge 30% more PRs with it, while people who primarily use Projects merge six times as many. You just need to chat with its coordinator agent.[Cursor Projects](https://cursor.com/blog/projects?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **OpenAI’s new voice model can listen while it talks.** GPT-Live-1 processes incoming and outgoing audio together, so voice agents can handle interruptions, pauses, and background speech without the awkward rhythm of turn-based systems. It costs $0.05 per minute for the voice layer, with the backend model and harness charged separately.[OpenAI’s announcement](https://openai.com/index/introducing-gpt-live-1-in-the-api/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Cognition released SWE-2, their most advanced coding model yet.** It scored 50% on FrontierCode 1.1, within one point of Fable 5.1, costing 64% lesser. Compared with SWE-1.7, its medium setting used 58% fewer turns, cost 81% less, and made its first meaningful edit after a median of 18 steps instead of 48. It is not just getting smarter, it is learning when to stop poking around and start fixing things.[Cognition’s SWE-2 report](https://cognition.com/blog/swe-2?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **AWS gives background agents the interface email figured out decades ago.** They open-sourced Pizza Bot, a self-hosted agent inbox where finished work lands in Unread and runs waiting for your approval move to Action. Tasks can start from a prompt, schedule, or webhook, survive closed tabs and interrupted sessions through on-disk checkpoints, and delegate to specialists built with MCP and Agent Skills. It is available now as a desktop app for macOS, Windows, and Linux.[Pizza Bot announcement](https://aws.amazon.com/blogs/opensource/introducing-pizza-bot-an-open-source-inbox-for-ai-agents-that-work-in-the-background/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) ## 🧠 **Worth Knowing** **Sebastian Raschka says GPT-6 Astra may be thinking in depth, not merely in tokens.** He separated what we know about GPT-6 Astra from the rumors, then explored one especially intriguing possibility: a looped transformer that repeatedly processes an internal state through shared layers. That could give the model more reasoning depth without producing pages of visible chain of thought.[Read Sebastian Raschka’s analysis](https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Anthropic built a calculator for the economic futures nobody can agree on.** Their scenario explorer turns your assumptions about AI capability, adoption, autonomy, and productivity into a picture of the US economy in 2030. In its extreme scenario, annual GDP growth reaches 15%, the economy doubles every 4.5 years, and unemployment climbs beyond ordinary recession levels.[Anthropic’s economic scenarios](https://www.anthropic.com/institute/econ-scenarios?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Matt Pocock is giving knowledge work a day shift and a night shift.** For course and talk planning, he creates a shared Karpathy-style LLM wiki for each deliverable, then splits the work into sections that separate agent threads can develop in parallel. During the day, he dictates large brain dumps while the agents keep the wiki updated. At night, a custom linting skill sends a fleet of subagents through the entire project to hunt for weaknesses.[Read Matt Pocock’s workflow](https://x.com/mattpocockuk/status/2097638166232457451?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Google scored thousands of agent submissions and found these four agent patterns that survived contact with real users.** Bidirectional MCP, event-driven concurrency, equally strict validation for fallback models, and cheap routing before expensive model calls. One team handled more than 40% of incoming requests with a zero-token regex layer. The best “multi-agent” systems, it turns out, are often distinguished by good engineering rather than more agents.[Google’s four engineering patterns](https://developers.googleblog.com/en/4-engineering-patterns-behind-the-strongest-ai-agents-challenge-submissions/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Pi tells you when your prompt cache quietly starts charging again.** Switching models, editing an earlier message, or leaving a session idle can cause a cache miss, forcing the model to reread context it has already processed. Pi can now detect these misses, notify you, and estimate their cost. It turns one of agent pricing’s most invisible leaks into something developers can finally see and debug.[Pi’s cache-miss announcement](https://x.com/pidotdev/status/2097648758888542527?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Magic says it matched DeepSeek V4 with 50x fewer FLOPs.** They claim its pretraining recipe then beat publicly available open base models with a larger $4 million run. The gains came from dozens of improvements across architecture, optimization, data, and training stability rather than one magic trick. The final model is not public and the results are self-reported, but if they hold, frontier pretraining may have far more algorithmic slack than giant compute budgets suggest.[Magic’s pretraining report](https://magic.dev/blog/pretraining?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **CursorBench got harder, and every model’s score dropped**. The new CursorBench 4.0 adds more difficult tasks that test instruction following and sustained work on challenging projects. Scores are lower across all models under the new benchmark, with Grok 4.6 taking a notable hit and Muse Spark performing comparatively well.[Explore CursorBench 3.2](https://cursor.com/cursorbench?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **The cyber kill chain is becoming an agent workflow.** Anthropic’s threat team published a report where they disrupted AI-assisted espionage, surveillance, fraud, influence campaigns, weapons work, biological misuse, and model theft. In one Russian-linked operation, agents monitored whether malware had been detected, modified and rebuilt it until security products stopped noticing, then staged it for live attacks. [Anthropic’s threat intelligence report](https://www.anthropic.com/threat-intelligence-report-september-2026?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) ## 🔧 **Clone and Run** **Clone & Run of the Day** **Stop guessing which models your laptop can survive.** llmfit inspects your CPU, RAM, GPU, VRAM, and accelerator setup, then ranks hundreds of open models by memory fit, expected speed, quality, and context capacity. You can also benchmark a model on your own machine and submit the real tokens-per-second result, replacing estimates for anyone with identical hardware.[llmfit](https://github.com/AlexsJones/llmfit?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Let your coding agent grab a real Android phone.** Google just dropped Artemis that lets Claude Code, Codex, Cursor, and other agents operate physical Android devices or emulators through MCP. It can execute cross-app workflows, reproduce bugs, collect Logcat crashes and screenshots, and continue exploratory testing for more than 10 hours. [Artemis](https://github.com/google/artemis?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Google Maps has packaged its API knowledge into installable agent skills** covering maps, places, geocoding, routing, Street View, air quality, weather, and more across web and mobile. The skills retrieve fresh documentation while generating code instead of trusting whatever API details the model remembers.[Google Maps agent skills](https://github.com/googlemaps/agent-skills?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Build an entire AI company on your own server.** OtoDock gives Claude Code, Codex, and local models persistent jobs, memories, workspaces, tools, schedules, departments, and even phone lines. Teams can share agents while keeping personal workspaces private, and every agent runs with an explicit set of knowledge, skills, and permissions. [OtoDock](https://github.com/OtoDock/oto-dock?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Turn the paper you have been avoiding into an animated lesson.** arXivisual converts an arXiv paper into a scrollable explanation with narration and Manim animations beside the concepts they illustrate. Change `arxiv.org` to `arxivisual.org` in a paper’s URL, and it breaks the paper into sections, chooses what to animate, writes the animation code, and narrates the result. [arXivisual](https://x.com/hasantoxr/status/2097398574061670664?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Give Claude Code and Codex one shared memory.** agent-memory stores long-term memories as plain Markdown, with a disposable SQLite index that can be rebuilt without losing knowledge. Claude Code, Codex CLI, and other agents can read the same store.[agent-memory](https://github.com/tigerless-labs/agent-memory?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **Find every AI agent that can reach your secrets.** Run `npx geiger-scan` and this tool inventories the agents, MCP servers, plugins, hooks, editor extensions, and AI browser extensions installed on your machine. It labels which ones can execute code, access broad parts of the filesystem, reach the web, or hold credentials, then points to the exact configuration that produced each finding.[Geiger](https://github.com/Atomburstofficial/geiger?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) [Awesome LLM Apps](https://github.com/Shubhamsaboo/awesome-llm-apps?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **is a curated collection of 100+ AI Agents, Agent skills, and RAG apps.** It covers models from OpenAI, Anthropic, Google, and open-source models like GLM, DeepSeek, and Qwen that you can run locally on your computer. [(Now accepting GitHub sponsorships)](https://sponsorunwindai.com/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) ## 📊 **By the Number** **Number of the Day****24 hidden bugs fixed for $1.08 with DeepSeek V4.1 Flash.** On Bug Hunt Bench, DeepSeek V4.1 Flash fixed 24 of 105 planted bugs across two repositories, versus 27 each for Opus 5 Max and Grok 4.6 Max. The updated max-effort run cost $1.08, compared with $51.33 for Opus and $16.96 for Grok, while its high-effort setting fixed 19 bugs for just $0.31. Getting 89% of the leaders’ result at a tiny fraction of the cost is insane![Bug Hunt Bench](https://bughunt.productcompass.pm/?preset=featured&utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) | [Read the complete test thread](https://x.com/PawelHuryn/status/2098002428054397185?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **A 3.8B model trained from scratch for $998.** One engineer trained little-lm, a 3.8B-parameter language model, on 65.3 billion tokens using eight rented B200 GPUs for 43 hours. The final bill was $998, and the model scored 0.384 on CORE, comfortably above GPT-2’s 0.2565 and a similarly priced nanochat run’s 0.310. The full write-up includes the architecture, training configuration, failed experiments, and optimizations that made the budget work.[How little-lm was trained](https://hugovergnes.github.io/little-lm-3-8b/?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) **98% of Astra’s design score at 1.4% of the cost.** In OpenDesign Arena’s evaluation of everyday design tasks, DeepSeek V4.1 Flash scored 81.2 out of 100, versus GPT-6 Astra’s 82.7, while costing $0.023 per artifact instead of $1.61. It finished in 5.3 minutes on average, less than half Astra’s 11.1 minutes, although results varied sharply by task: DeepSeek beat Astra on landing pages but trailed it on dashboards.[OpenDesign’s results thread](https://x.com/OpenDesignHQ/status/2097620919778955617?utm_source=www.theunwindai.com&utm_medium=referral&utm_campaign=openai-codex-harness-as-an-api) That's all for today. Come back tomorrow for the next batch of AI tools, model drops, agent repos, and weird benchmarks worth your time. If you found one thing to try, share the issue with someone who ships. ### Hire Ava, the AI BDR built for enterprise Ava is the [first AI BDR](https://www.artisan.co/request-demo?utm_source=beehiiv&utm_medium=newsletter&utm_campaign=enterprise_demo&utm_content=JHL0VVEUDT&_bhiiv=opp_255994c3-f602-4b70-8777-506389c20948_f504ea8d&bhcl_id=9032ab6e-1626-4e17-a38d-179e75d70843_SUBSCRIBER_ID_{{email_address_id}}) to run outbound end to end, and you decide whether she runs autonomously or on copilot. She finds leads or ingests accounts from your CRM, enriches them, sends personalized emails on behalf of your reps, follows up, handles replies and books meetings. Website visitor de-anonymization, intent signals, and a parallel dialer come built in. A small team can manage her centrally for thousands of reps who never log in. Everything syncs two-way with Salesforce and HubSpot. [Ava](https://www.artisan.co/request-demo?utm_source=beehiiv&utm_medium=newsletter&utm_campaign=enterprise_demo&utm_content=JHL0VVEUDT&_bhiiv=opp_255994c3-f602-4b70-8777-506389c20948_f504ea8d&bhcl_id=9032ab6e-1626-4e17-a38d-179e75d70843_SUBSCRIBER_ID_{{email_address_id}}) runs outbound for companies like DoorDash and Grammarly, and one customer deploys her across 1,000+ reps. Ava is SOC 2 Type II audited, SSO and GDPR ready. Ava is how revenue teams grow pipeline without growing headcount.