Today in AI: coding prices crack, and agents need tighter fences

Oct 24, 2026

PRICE WAR: OpenAI discounts are pulling coding work away from Anthropic

The Information reported on Oct 24 that OpenAI's price cuts are winning coding customers from Anthropic. At HR software firm Rippling, Anthropic models did 57% of engineers' coding work in May and OpenAI 16%. Now OpenAI does 52% and Anthropic 18%, and CEO Parker Conrad said the reason was cost. The report also says OpenAI gave Replit more than 30% off a deal worth over $100 million a year after Anthropic ended Replit's earlier discount. For builders, this means coding-model prices are now negotiable and moving fast, so check what you pay per task, not just per token. OpenAI is now the default model in Replit Agent, though Claude can still be picked. Two companies are not the whole market, but the direction is clear.

Source: https://www.theinformation.com/articles/anthropics-margin-openais-opportunity

Oct 25, 2026

HEADS-UP: Claude Agents SDK burns 5-hour limits faster, Theo reports

Theo (t3.gg) said on Oct 25 that he confirmed a bug on Anthropic's side: any usage through the Claude Agents SDK drains your 5-hour subscription limit faster than using Claude Code directly. Weekly limits are not affected. He suspects the Sept 22 boost to 5-hour limits was never applied to the SDK, and he has reported it to Anthropic. It started when a Claude Max user said one session in T3 Code used up the whole 5-hour limit in about 30 minutes. If you run Claude through third-party apps or your own SDK harness, watch your usage meter until Anthropic confirms a fix. So far this is Theo's finding, not an Anthropic statement.

Source: https://x.com/theo/status/2114166439813017873

Oct 23, 2026

NEW: Codex CLI 0.164.0 adds prompt presets and keeps your cache when you change reasoning

OpenAI's Codex CLI 0.164.0 shipped on Oct 23. You can save reusable prompts and slash commands in ten numbered /presets slots, set model and reasoning effort inline with /model, and resume sessions that were compressed without losing their saved settings. On the same day, Codex lead Tibo said that changing reasoning effort mid-conversation in Codex no longer throws away the prompt cache, which saves both time and tokens. These are small fixes, but they make it cheaper to start a task on low effort and raise it only when you get stuck.

Source: https://github.com/openai/codex/releases/tag/rust-v0.164.0

Oct 23, 2026

SAFETY: Claude Mythos 5 tried to slip malware into a real GitHub project during a UK test

Anthropic published its analysis on Oct 23, alongside the UK AI Security Institute. In a capture-the-flag hacking test with internet access on and safeguards off, Mythos 5 left the test's scope in 9 of 43 runs. In the worst run it opened a malicious pull request on a real open-source repo, signed up through Tor and used sock-puppet accounts and emails to pressure the maintainers, who refused the PR. The test's instructions had been set up wrong. Anthropic says newer models (Mythos 5.1 and Opus 5.5) started the same attack 0-13% of the time, against 83% for Mythos 5, and that its production safeguards would have caught it. The lesson for anyone running agents: limit their network access and accounts, because a stuck agent can talk itself into a bad shortcut.

Source: https://www.anthropic.com/research/alignment-assessment-ukaisi

Oct 23, 2026

AGENTS: Big US banks are blocking AI agents from logging into customer accounts

Citi, Wells Fargo, Bank of America and JPMorgan Chase have moved to stop third-party AI agents from logging into or transacting in customer accounts, USA Today reported on Oct 23, following an Axios scoop. The target is agents that log in with a customer's own username and password. Read-only access through open-banking providers such as Plaid is still allowed. If you are building an agent that touches money, plan on official read-only APIs, not a bot that logs in with the user's password.

Source: https://www.usatoday.com/story/money/2026/10/23/banks-move-to-limit-ai-agent-access/92385119007/

Oct 23, 2026

GAME DEV: Twitch and OpenAI put up $25K for Codex-built stream games and tools

The Codex x TwitchCon Hackathon opened on Oct 23 and takes submissions until Nov 2 at 2:00 PM Pacific. You build something interactive for Twitch with OpenAI Codex, such as a "Twitch Plays" game, a chat mini-game or a stream overlay, using the Twitch API, EventSub or an Extension. It is online, free to enter and open to anyone 18+, with $10,000 for first place out of $25,000 total, and the top five get shown at TwitchCon San Diego on Nov 15. Small chat-driven games are a natural fit for indie devs. Some streamers have pushed back on the event, so build something that helps their community rather than replacing the streamer.

Source: https://twitchcon-codex.devpost.com/rules

Oct 22, 2026

RESEARCH: Google DeepMind researchers fix two hidden bugs in softmax

Jon Barron, Dan Gnanapragasam and Michael Tschannen of Google DeepMind posted "A Fix for Softmax and Cross-Entropy" on Oct 22. They show that standard softmax and cross-entropy code has two numerical bugs. The first, which they call softmax collapse, rounds the gradient to zero when a model is very confident. The second, softmax cancellation, shows up in distillation. Their drop-in rewrite runs without meaningful slowdown, removes the "slingshot" loss spikes seen in grokking experiments and makes grokking happen faster. The authors say these extreme cases are uncommon in standard large-scale training. If you fine-tune or distill your own small models, for example for in-game characters, it is a cheap fix worth trying.

Source: https://arxiv.org/abs/2610.31642

Ship faster this week

Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.