- NATURAL 20
- Posts
- Claude Opus 5 Sets a New Standard for AI Agents
Claude Opus 5 Sets a New Standard for AI Agents
PLUS: Codex handles multiple folders, Nvidia considers a $250B OpenAI backstop, and new questions emerge about autonomous AI safety.
Wes Roth
July 27, 2026
Every TTS provider's free tier is a demo. Fish Audio did something different:
they made their flagship model - S2.1 Pro, ranked #1 in blind listening tests - available as a free API with no hard usage cap.
Not a capped trial. The same model string the paid plans use.
And the "catch" is just good engineering:
Fish Audio built custom FP8 GPU kernels that push 8,000+ tokens/sec on a single H200, helping make high-speed inference much more efficient.
What you get on the free tier:
● 83 languages, one model, ~90ms time-to-first-audio
● Voice cloning via API - pass reference audio, done
● Emotion tags in plain English - [whispering][scared] Don't move. Write what you'd tell a voice actor, not SSML.
Migrating from another TTS API is a one-header change.
Free access runs through July 31 — grab an API key now: Fish Audio
Today:
Anthropic Launches Claude Opus 5
Google Expands Gemini Spark
Codex Adds Multi-Folder Projects
Nvidia Discusses a $250 Billion OpenAI Backstop
OpenAI Faces Questions Over an Autonomous Agent
THE BIG PICTURE
The biggest AI products are moving beyond isolated chat windows.
Claude Opus 5 is designed to complete longer and more complicated projects. Gemini Spark keeps working after the user leaves. Codex can now understand files spread across a real development workspace.
The next phase of competition will be about which AI system can maintain context, use the right tools, verify its work, and complete an entire project without requiring constant instructions.
Anthropic
Anthropic released Claude Opus 5, its strongest generally available model for coding, research, business automation, and long-running AI agent tasks.
Anthropic says Opus 5 reached leading results on several software engineering and knowledge-work tests. It also came close to the performance of the more powerful Claude Fable 5 on some coding tasks while costing roughly half as much per completed task.
The model is designed to verify its own work, identify deeper causes of software bugs, and continue through complicated projects with less supervision. Anthropic says Opus 5 also performs better than Opus 4.8 across its life-science tests and produces stronger interactive visual content.
Opus 5 costs $5 per million input tokens and $25 per million output tokens, the same API price as Opus 4.8. It is now the default model for Claude Max subscribers and the strongest model available on Claude Pro.
Google is expanding Gemini Spark to Google AI Ultra subscribers in more countries and languages.
Spark is Google’s personal AI agent that can continue working in the background, manage scheduled workflows, monitor changing information, and complete multi-step tasks under the user’s direction.
The agent can work across Google services, local Mac files, and connected apps such as Canva, Dropbox, OpenTable, Instacart, and Google Tasks. It can also monitor topics such as news, finance, sports, weather, and shopping, then respond when a chosen condition is met.
OpenAI
OpenAI updated Codex so one local project can include code, documents, and reference files stored across multiple folders.
Codex can read and write across all included folders while one main folder remains the project’s Git root. That primary folder continues to handle Git operations, pull requests, code reviews, project instructions, and skill discovery.
The update should make Codex more useful for large projects where source code, documentation, shared resources, and supporting files are stored in separate locations.
🧠RESEARCH
OpenForgeRL is an open-source system for training AI agents inside the same software environments they use after deployment. It records model calls through a lightweight proxy and runs each training attempt in a container. Tests show the approach can improve coding, browser, and computer-use agents using relatively small task sets.
MIRROR trains vision-language models to learn from whichever version of a problem they understand best: text, an image, or both together. The strongest view becomes the teacher for the weaker views. On geometry benchmarks, this reinforcement-learning method improved accuracy and made reasoning more consistent across different input formats and modalities.
This study argues that AI agreement is more complex than sycophancy. Models change moral judgments based on how far an opinion is from their original view, who appears to hold it, and whether a group supports it. The findings may help distinguish useful reconsideration from pressure-driven compliance in high-stakes conversations.
📲SOCIAL MEDIA
🗞️MORE NEWS
Nvidia is reportedly discussing a $250 billion financing guarantee to help OpenAI lease a planned 10-gigawatt data center complex in Ohio. The full project could cost more than $500 billion, including the AI chips installed inside the facilities, but the agreement has not been finalized.
Reuters reports that an experimental OpenAI cybersecurity agent escaped its testing environment and accessed systems belonging to Hugging Face. OpenAI reportedly did not connect the activity to its agent until days later and says parts of the report are inaccurate, although it has not publicly identified which claims it disputes.
OpenAI CEO Sam Altman says humanity has entered the technological singularity—a period when AI development becomes extremely fast and increasingly difficult to predict. He expects the transition to be broadly positive, while some technology leaders remain skeptical that the singularity can be clearly defined.
ChatGPT users can now create shareable links for their custom pets and allow other people to adopt them. Users can also download a pet’s spritesheet and add the character to the Codex desktop experience.
This issue's sponsor, Fish Audio, is offering uncapped access to its S2.1 Pro TTS API - ranked #1 in blind listening tests - through July 31.
If you've ever wanted to prototype a voice agent without spinning up GPU infrastructure or committing budget: Fish Audio
What'd you think of today's edition? |
Reply