Coding agents

OpenAI wins Anthropic customers on price, Block's Dessa, AWS Infinity, Ollama speaks: the October 25 roundup

OpenAI lures Replit and Rippling with discounts, Block unveils Dessa, AWS open-sources the Infinity runtime, Ollama speaks and Qwen Code reaches 0.26.1.

By Audryk · October 25, 2026

• Coding agents
• Models
• Cloud & infrastructure
• #OpenAI
• #Anthropic
• #pricing
• #agents
• #open source
• #Ollama

Sunday is quiet for announcements, so this roundup covers October 23 and 24, 2026, with five stories our previous editions had not covered: a press investigation into the price war between OpenAI and Anthropic, a voice model from Block, an agent runtime from AWS and two open source tool updates. Unless stated otherwise, figures come from official sources.

OpenAI wins Anthropic customers on price (Oct 24)

US outlet The Information reports that OpenAI is pulling major customers away from Anthropic with aggressive discounts (The Information, paywalled, summarised by The Decoder). This is a press investigation, not a company announcement.

• Replit: according to an anonymous source cited by The Information, the maker of Replit Agent moved most of its AI spending to OpenAI, with a discount of more than 30% on a contract worth more than $100 million a year, after Anthropic ended the discount it had been giving Replit.
• Rippling: CEO Parker Conrad says OpenAI's share of the company's coding work rose from 16% in May to 52%, while Anthropic's fell from 57% to 18%.
• DataCamp: CEO Jonathan Cornelissen says OpenAI's pricing cut his costs by 80 to 90%.

To compare the public prices of both model families, see the Claude and GPT pages and our cost calculator.

Block unveils Dessa, a voice model trained only on speech (Oct 23)

Block's AI lab (Square, Cash App) presents Dessa, a conversational voice model released as a research preview (Block's page). It was announced on X on October 23 by Block and its founder Jack Dorsey, then covered on October 24 by TechCrunch.

• How it works: a speech-to-speech, full-duplex model that listens and talks at the same time. Block says it trained it from scratch on human conversation, with no text and no pretrained model.
• Resources: training fit on a single NVIDIA GB200 NVL72 system, and inference runs locally on a MacBook Pro M4 Max or better, according to Block.
• Availability: no public access yet, only a waitlist. Block is giving early access to a small group of AI safety researchers.

AWS open-sources Infinity, a runtime for tens of thousands of agents (Oct 23)

AWS releases Infinity under the Apache 2.0 license, a research project built with UC Berkeley as part of the Hydro project (AWS Open Source Blog, GitHub repository).

• The idea: a tool call becomes a suspension point, like an await. The runtime saves the conversation, frees memory and resumes the agent when the result arrives, even days later.
• Reported figures: about 100 KB of memory per idle agent, or roughly 75,000 agents in the 8 GB of a Raspberry Pi 5. The benchmark ran 80,000 agents and 3.2 million completions on a single thread.
• Serverless: a reference deployment uses SQS FIFO, Lambda, Aurora DSQL and DynamoDB. Asynchronous events reach the model as synthetic tool calls (receive_event__injected). AWS describes it as an early research release.

Ollama 0.40.11: speech output and decision models (Oct 24)

Ollama 0.40.11 (release notes) adds two features. To install Ollama and use it with an agent, see our Ollama guide.

• Speech on macOS: ollama run can read text or a chat model's reply aloud, and apps can stream the audio through the OpenAI-compatible /v1/audio/speech endpoint.
• d3 models: vLLM Semantic Router's family of decision models, in 2B, 4B, 9B and 27B sizes, to classify, answer yes or no, or score options with probabilities.

Releases

• Qwen Code v0.26.1 (Oct 24): opt-in persistence of the daemon token (qwen daemon start --save-token), idle-only manual context compression for managed sessions, scheduled tasks gated on a shell command run outside the model; one breaking change affects the QQ Bot integration (release notes).
• Ollama v0.40.10 (Oct 23): adds NVIDIA's Nemotron 3.5 Bolt E4B, a small Gemma 4-based vision-language model with a 128K-token context window (release notes).

The number: 100 KB per waiting agent

That is the memory AWS measured for an idle agent on Infinity, mostly its conversation history (AWS). An agent waiting on a build or an approval no longer holds a process.

What it means

Cost is becoming as strong an argument as quality: according to the press, coding customers switch providers over a discount. At the same time, infrastructure is getting ready for many patient agents, and local tools such as Ollama are expanding into voice.

Sources

• The Information: Anthropic's Margin is OpenAI's Opportunity (paywalled, verified 2026-10-25)
• The Decoder: OpenAI offers discounts up to 90 percent to poach Anthropic's biggest customers (verified 2026-10-25)
• Block: Dessa (verified 2026-10-25)
• TechCrunch: Block's new voice model sounds eerily human (verified 2026-10-25)
• AWS: Introducing Infinity, a runtime for extreme agent concurrency (verified 2026-10-25)
• Ollama: v0.40.11 release notes (verified 2026-10-25)
• Ollama: v0.40.10 release notes (verified 2026-10-25)
• Qwen Code: v0.26.1 release notes (verified 2026-10-25)

Spotted an error in this article? Report it · Our corrections policy