GenAI Daily - October 9, 2026: Google's Gemini Agent, Ecosia Drops Mistral, Manus Raises $500M
Top Stories
Google Ships a Single "Gemini Agent" for Work, Routing Across Gemini and Claude
At Gemini at Work 2026, Google Cloud announced the Gemini agent, a "new single, universal agent for work" available in the Gemini Enterprise app, Workspace, and third-party services.
It runs across web, mobile, desktop, command line, Google Workspace, Microsoft 365 and Slack, and can also run headless inside third-party apps.
Teams can create a coworker agent with its own email address, calendar, Drive access and place in the corporate directory.
Colleagues can tag it in a Chat space or a Docs comment, and it only sees what is shared with it.
The engineering details are in the model routing and cost controls. It supports Gemini and Anthropic's Claude models, with more planned, and a Smart Routing feature assigns each workload to the model with the best quality-to-cost balance.
Project-level spending limits pause agent activity when budgets run out, and costs can be attributed to individual departments.
Each agent has a cryptographically attested identity with minimal permissions approved by security administrators.
Availability is limited: it is in private preview with no confirmed release date.
Pricing reports conflict. Google told VentureBeat it will be included at no additional charge where Gemini Enterprise is available. Other outlets report that no price was given in the keynote.
Google Cloud blog | 9to5Google | VentureBeat
Why it matters: Google is putting model choice, spend caps and agent identity inside one product, and it is routing to a competitor's models to do it, which raises the bar for governance features in any agent platform you evaluate.
UPDATE: Ecosia Drops Mistral for Chinese Open-Weight Models as Large 4 Gets Independent Scores
Berlin-based search engine Ecosia is replacing Mistral as its AI supplier. It had swapped OpenAI for Mistral only in May.
CEO Christian Kroll told Politico that Ecosia "was simply too large a customer for Mistral" after recurring server overloads, and he called the models "a year behind" the competition. Ecosia now gets its models through Melious, a German platform running open models on EU servers, with a lineup that includes Z.ai's GLM 5.3, Moonshot's Kimi K3 and DeepSeek V4.1 Flash.
Kroll says costs roughly halved. That figure comes from the CEO and has not been independently verified.
The timing is awkward for Mistral, whose Large 4 preview arrived this week with open weights due by month end. Artificial Analysis scores it 38, the strongest open model outside China but behind seven Chinese models.
On the Vals Index it placed 32nd of 44 models, though it took first place among open-weight models on Harvey's legal agent benchmark.
Mistral calls its own figures preliminary and expects them to shift before October 27.
Politico via AI Weekly | PPC Land | Trending Topics
Why it matters: For teams buying on a "European sovereignty" basis, capacity limits and a quality gap against Chinese open-weight models are now documented procurement risks, so benchmark on your own workloads before committing.
Workday Bets on Governing the Outside Agents That Query Its Data
Agents built on Claude or Microsoft's Copilot Studio can now read and act on HR and finance records in Workday without touching Workday's own AI. Anthropic and Microsoft agents can query Workday data directly, skipping its own AI tools.
The numbers so far don't show Workday being sidelined: over 5,500 customers were using its AI agents as of Q2 2026, up more than 35% quarter over quarter.
Workday's response is to govern outside agents rather than block them. Its Agent System of Record works as a registry for Workday's and everyone else's agents, with security controls, permission handling and audit trails.
External agents come in through Agent Gateway using MCP and A2A, and direct calls to Workday APIs are now metered per call.
This contrasts with Salesforce's approach. At Dreamforce, AIforce put Salesforce data, workflows and permissions inside Claude, Slack and other AI tools via MCP, APIs, plugins and skills. Both vendors are conceding the interface to third-party agents and trying to keep the data, policy and metering layer.
Why it matters: If you plan to point Claude or Copilot agents at Workday, budget for per-call API metering and registration through the gateway, because the vendor's licensing leverage now sits there.
Key Developments
OpenAI's 722-Manuscript Math Release Draws a Boycott Call and a Retraction
On October 6, OpenAI opened a repository of 722 manuscripts grouped into 372 families of results, plus a Lean library.
OpenAI had to retract three papers the next day over a sign error. The Association for Human Mathematics responded with a statement that Terence Tao reposted. It urges mathematicians to discontinue their work with OpenAI.
OpenAI says nearly all 372 results came from a single prompt to one agent, but it withheld the prompts and per-problem compute times despite advisory-group requests for full disclosure.
Terence Tao's blog | The Decoder
Impact: The dispute is about verification and cost transparency, and the same gap exists in vendor agent benchmarks, so ask for compute and retry data before trusting headline results.
Adobe Puts Its CX Agent Inside ChatGPT and Lets Customers Pay With OpenAI Spend
Adobe launched its CX Enterprise Coworker plugin in ChatGPT alongside OpenAI DevDay. It brings audience curation, journey design, asset creation, offer decisioning and channel activation into one persistent session.
Adobe is also a launch partner for OpenAI's Marketplace, so customers can apply existing OpenAI spend toward a license, and an Adobe Workfront integration with OpenAI's Agents API is scheduled for October.
Impact: Marketplace billing means agent purchases can come out of an existing OpenAI commit, which changes how procurement and budget owners handle these tools.
Product Launches
Google Gemini 3.8 Flash Introductory Pricing Ends December 31
Google's Agent Platform pricing page lists Gemini 3.8 Flash at introductory rates of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026.
Starting January 1, 2027, input rises to $1.50 per million.
Impact: Teams with high-volume workloads should model the 2x input increase now.
Docker Agent v1.149
Docker's Apache-2.0 agent runtime defines agents and multi-agent teams in YAML and runs them with docker agent run. It supports OpenAI, Anthropic, Gemini, Bedrock, Mistral, xAI and Docker Model Runner, uses MCP natively, and can push and pull agents from any OCI registry.
Docker Desktop 4.63 and later already include it.
Impact: Existing registry, scanning and CI workflows can cover agent definitions.
Funding & Deals
Manus Raises $500M+ After Meta Deal Unwinds
Manus, the Singapore-headquartered developer of general-purpose AI agents, closed its first round since Beijing ordered Meta to unwind its $2 billion-plus acquisition in April.
The company did not disclose a valuation, though Bloomberg previously reported a $4 billion target.
The Information reported an annualized revenue run rate of about $500 million, up from $100 million at the time of the Meta deal.
A source says a Hong Kong IPO process will not start until at least 2027. Led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund participating.
Arena Raises $200M Series B at $3.1B Valuation
Arena, the company behind the LMArena leaderboard, started in 2023 as a UC Berkeley research project crowdsourcing model rankings.
It is now measuring models on alignment issues such as lying, and it reached $100 million in annualized run-rate revenue in June. Led by Lightspeed and Khosla Ventures, with Salesforce Ventures, Dell Technologies Capital and a16z joining.
Stuut Raises $52.5M Series B
New York-based Stuut sells AI agents that run order-to-cash, covering order management, credit, collections, payments, cash application, disputes and deductions.
It integrates with ERPs such as SAP, Oracle, NetSuite and Microsoft Dynamics and keeps audit trails and financial controls.
More than 150 companies use it, including Fortune 50 and 500 businesses, and it reports customers cutting days sales outstanding by 47%. Those figures are company-reported.
Total funding is now $93 million. Led by Insight Partners.
Tomorrow's Watch List
- Anthropic IPO: a follow-up report says it will meet prospective investors on October 14, and marketing could begin the week of November 9.
- TechCrunch Disrupt runs October 13-15 at Moscone Center in San Francisco.
- Salesforce October GA targets: AI Skills in Agentforce Coworker and Agent Optimizer are both slated for general availability in October.
- Mistral Large 4 open weights are due October 27, which will allow independent reruns of the preview benchmarks.
*Related reading: Check out this week's [Deep Insights analysis] for strategic context on these developments.
