In a single month, Anthropic, OpenAI and Google all shipped new models built for agentic work: AI that does not just answer, but clicks, types, calls tools and finishes tasks. Here is what launched, and what it means if you run a business rather than an AI lab.
Anthropic: the Claude 5.1 and 5.5 generation
On 1 September Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the same model with two levels of safeguards. Fable 5.1 is generally available; Mythos 5.1 is restricted to vetted professionals in cybersecurity and life sciences. Alongside it came Enterprise Frontier Safeguards, which let enterprise customers keep their data on their own cloud infrastructure, rolling out in phases from this autumn.
Three weeks later, on 22 September, Claude Opus 5.5 arrived as the first model of the Claude 5.5 family. Anthropic says it performs at the level of Fable 5.1 on most work and costs about 40% less to run than Opus 5, at $4 per million input tokens and $20 per million output tokens. It reports 81.8% on the OSWorld 2.0 computer-use benchmark.
OpenAI: GPT-6 Astra, Sol and Luna
OpenAI positions GPT-6 Astra as its most capable model, with computer use that can fill in online forms, update CRM records and organise calendars. OpenAI reports 72.6% on OSWorld 2.0; API pricing is $10 per million input tokens and $50 per million output tokens.
On 22 September OpenAI added GPT-6 Sol and GPT-6 Luna as cheaper members of the family. Sol costs $2 / $10 per million tokens and, according to OpenAI, reaches computer-use scores comparable to Claude Opus 5 at roughly 80% lower cost per task. Luna costs $0.10 / $0.50, cheap enough to run high-volume classification and routing.
Google: Gemini 3.8 Flash
On 2 September Google launched Gemini 3.8 Flash, described as its most intelligent workhorse model, with gains in software engineering, agentic tasks and multi-step reasoning. Introductory pricing is $0.75 / $3.75 per million tokens until 31 December 2026, then $1.50 / $7.50. A separate Gemini 3.8 Flash Cyber model for vulnerability detection is limited to authorised security professionals.
Open standards: MCP and A2A under one roof
On 27 August the Agent2Agent (A2A) protocol joined the Linux Foundation's Agentic AI Foundation, next to the Model Context Protocol (MCP), goose and AGENTS.md. MCP connects an agent to your tools and data; A2A lets agents from different vendors discover each other and delegate work. More than 150 organisations back A2A, including Google Cloud, AWS, Microsoft, Salesforce, SAP and ServiceNow.
What this means for businesses
- Computer use is becoming practical. Agents can now operate web apps and back-office tools that have no API, which opens automation for legacy systems.
- Cost per task is falling fast. Mid-tier models such as GPT-6 Sol, Gemini 3.8 Flash and Claude Opus 5.5 deliver near-frontier results at a fraction of last year's price.
- Use a model ladder. Route simple steps to the cheapest model that passes your tests and escalate only the hard steps to frontier models.
- Security and data residency are now product features. Restricted models and customer-hosted data options make regulated use cases easier to approve.
- Benchmarks are vendor-reported. Always test on your own data and workflows before you commit.
Where to start
The fastest way to turn these releases into value is a short, measured pilot on one real workflow. DanaExperts runs paid two-week AI pilots that connect the right model to your systems and report accuracy, speed and cost, so you decide on numbers rather than headlines.
