Last checked October 18, 2026 · News roundup
Today in brief: Replit has made OpenAI its default model provider for Replit Agent, a public split from Anthropic that its CEO tied to cost and performance. DeepSeek shipped a new V4.1-Flash checkpoint with paid Search and Fetch APIs, and Anthropic published hard numbers on how AI agents are reshaping its own hiring. On the business side, OpenCode told some users to rotate their AI provider keys after a breach, Meta walked back the AI-training wording in its business terms, and AI software stocks had a rough Friday.
This is a weekend edition, so the list leans on stories that broke late Friday and Saturday. Each item ends with a practical next step. If you only have a minute, read the OpenCode section: it is the one that can cost you money right now.
Part 1 · AI
Replit makes OpenAI its default model: what changes for Replit Agent users
Replit CEO Amjad Masad said on stage at TechCrunch Disrupt that the company is "no longer working with Anthropic." TechCrunch, which first reported the remarks on October 16 and updated its story on October 17, says Masad then clarified the scope: OpenAI is now the default model provider for Replit Agent, and customers can still pick Anthropic's Claude models in the product.
His stated reasons were "cost, partnership, and performance," adding "I think ChatGPT is better now" for building vibe-coded apps from scratch, based on Replit's internal and public benchmarks. Masad said Replit had been a top-five Anthropic customer until recently. A person familiar with the partnership told TechCrunch the arrangement is still in place, and Anthropic has not commented publicly.
Who it affects: anyone who builds with Replit Agent and leaves the model setting alone, which TechCrunch notes is most users. No price change was announced, but the style and speed of Agent output can shift when the model underneath changes.
What to do: if a project depends on Claude's behavior, check the model picker before your next long Agent session and pin the model you want. Our Replit review covers how the Agent, deployments and plans fit together, and the AI app builders hub lists alternatives if you would rather keep a Claude-first workflow.
Is Claude Fable 5.5 out yet? What Anthropic has confirmed so far
No. As of October 18, Anthropic has not announced a model called Claude Fable 5.5, and the newest Fable in its documentation is still Fable 5.1. The rumors come from a string of leaks.
The evidence so far, all unconfirmed by Anthropic: TestingCatalog reported Fable 5.5 traces in Azure configurations on October 13, and a screenshot circulated showing a claude-fable-5-5 listing in Azure's East US 2 catalog marked "Generally available." A LiteLLM model entry for it, with pricing that matches Fable 5.1, was removed about a day later. TestingCatalog's October 17 report also cites leakers who say partners have access and expect a release next week.
What it costs: nothing official. The LiteLLM entry showed $10 per million input tokens and $50 per million output tokens, which is Fable 5.1's list price, so treat it as a placeholder rather than a leak of new pricing.
What to do: don't hard-code a model ID you saw in a screenshot. If you run Claude through Azure, Bedrock or a router such as LiteLLM, wait for Anthropic's own model page or release notes before switching traffic.
DeepSeek V4.1-Flash-1017: new checkpoint and a $0.75 search API
DeepSeek released V4.1-Flash-1017 on October 17, a new checkpoint of its fast model that is served on its website, app and API under the existing deepseek-flash name. According to MarkTechPost's write-up, the architecture pairs a 552B backbone with 196B "Engram" parameters, activating about 8B per token during prefill and 16B during decoding, with a 1M-token context window. DeepSeek's model card shows the new checkpoint improving on the September Flash on every listed benchmark, for example 77.0 vs 74.2 on DeepSWE v1.1.
The bigger change for developers is billing. DeepSeek now sells server-side web Search and Fetch tools, each at $0.75 per 1,000 successful requests. Empty results, failures and timeouts are not charged, and each tool is capped at 5 requests per second per account. Calls made through DeepSeek Harness, Claude Code or Codex are free until billing starts at 00:00 Beijing time on November 1, 2026.
DeepSeek API prices after the October 17 update
Input, cache miss
$0.30 / 1M
Peak hours. $0.15 off-peak.
Output
$1.20 / 1M
Peak hours. $0.60 off-peak.
Search or Fetch
$0.75 / 1,000
Successful calls only. 5 requests per second per tool.
Free window
Until Nov 1
For calls routed through DeepSeek Harness, Claude Code or Codex.
Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays; weekends are off-peak.
What to do: if you plan to use DeepSeek's search inside Claude Code or Codex, set a usage alert before November 1 so the switch from free to paid does not surprise you.
Anthropic employment report: junior hires fell from 27% to 9%
Anthropic published "How employment is changing at Anthropic" on October 16, a rare look at what heavy internal use of AI agents does to a company's own workforce. The junior share of technical staff dropped from 21% to 14% since early 2025, and the junior share of technical hires fell from 27% to 9%. By August, interactive hours logged by agents that employees launched outnumbered employee hours six to one.
Anthropic's workforce mix, before vs. now
Early 2025 / a year agoLatest figure
Junior share of technical staff
21%
14%
Junior share of technical hires
27%
9%
Technical departments' share of all staff
61%
56%
Other numbers in the report: code commits per engineer are about five times what they were six months earlier, median meeting time rose 20% year over year, and roughly 80% of interactive Claude session time is spent waiting for a human. The authors note that agent adoption was not random, so the productivity estimates may be biased, and they found no rise in deletion or reversion rates for agent-written code.
Why it matters to you: hiring managers can treat this as one company's data, not an industry forecast. It does show where the pressure lands first: entry-level technical roles and anyone whose work is mostly writing first drafts of code.
Anthropic's first Canadian data center: 95 MW in Nova Scotia by late 2027
Anthropic signed a long-term lease with EverWind Fuels for a data center campus at Point Tupper, Nova Scotia, Reuters reported on October 16. It is Anthropic's first data center deal in Canada.
According to TokenPost and Crypto Briefing, the lease covers phase one, 95 megawatts targeted for late 2027, with a term of at least 20 years. A second 325 MW phase planned for 2028 would take the campus to 420 MW, subject to grid studies and approvals. EverWind says it will invest more than C$10 billion, about C$8 billion for the data center and C$2 billion for onshore wind, with roughly 2,000 construction jobs and about 215 permanent roles. Reuters adds that Anthropic reported $7.33 billion in computing and infrastructure spending last year.
Who it affects: not Claude subscribers directly. New capacity mostly matters over time, through fewer rate limits and outages, and for Canadian enterprise and public-sector buyers who ask where their data is processed. Nothing changes in Claude plans or prices today.
Part 2 · SaaS & business
OpenCode Console breach: who needs to rotate API keys now
OpenCode, the open-source coding agent that lets developers plug in their own model provider keys, has emailed affected customers about a security breach, Notebookcheck reported on October 17. The company said an attacker gained unauthorized access to "a legacy system it was in the process of decommissioning," and linked the incident to what it called a large-scale distillation attack.
The exposure is limited to 1.4% of OpenCode Console users, and only accounts that had linked external providers. Third-party keys stored in those workspaces, reportedly including OpenAI and Anthropic keys, may have been taken. OpenCode says users who did not get an email are not affected, that the incident is ongoing, and that it is working with Google, OpenAI and Anthropic. Some users have posted unexpected charges, including one claim of $400 on Anthropic; those reports are unverified, and OpenCode has not said whether it will cover them.
If you got OpenCode's breach email: do these four things
Revoke every provider key you stored in OpenCode Console (OpenAI, Anthropic, Google) from each provider's own dashboard.
Create new keys and update every app, CI job and script that used the old ones.
Check usage and billing pages for the last few days and screenshot anything you did not run.
Set spend limits or usage alerts at each provider, then contact their support about disputed charges.
Meta business terms update: AI-training crawl language removed
This is a follow-up to the change we flagged in Friday's roundup. Meta's revised Commercial Terms had said it could use content from the domains and web pages linked to a business account to develop and train its products, which publishers read as permission to train AI on their sites.
On October 17, TechCrunch, citing The Verge, reported that the provision was removed after pushback. Meta's statement: "the crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves," adding that the update "does not change how Meta uses web crawlers." According to Unite.AI, the revised terms still take effect on October 30.
What to do: if your company paused linking a website to its Meta business assets over this clause, re-read the current version before October 30.
Why AI software stocks fell on October 16
The Nasdaq Composite dropped about 2.15% to 26,140 on Friday, October 16, while the S&P 500 fell about 1.07% to 7,661 and the Dow slipped 0.34% to 52,002, according to a Motley Fool market wrap on Yahoo Finance. Palantir fell 7.22% and entered correction territory, and AI cloud provider CoreWeave finished the week down 25%.
Friday's damage, in percent (Oct 16, 2026)
Dow Jones: −0.34%
S&P 500: −1.07%
Nasdaq Composite: −2.15%
Palantir: −7.22%
CoreWeave, full week: −25%
The same report cites Goldman Sachs saying hedge funds sold AI stocks at the second-fastest pace in more than a year, which Goldman described as a "tactical rotation" from smaller AI names toward steadier large AI names such as the Magnificent Seven, and in some cases into non-AI stocks. For software buyers, a sell-off like this tends to show up later as tighter budgets at smaller AI vendors, so it is a reasonable moment to check the cancellation and data-export terms of any young AI tool your team relies on. This is market context, not investment advice.
San Francisco data center moratorium: a 45-day pause, signed October 16
San Francisco Mayor Daniel Lurie signed a citywide pause on new data center development on Friday, October 16, the San Francisco Standard reports. The measure, written by Supervisor Shamann Walton, passed the Board of Supervisors unanimously the week before and would have taken effect even without his signature.
The pause lasts 45 days while the Planning Department studies the impact of data centers and drafts rules for them, and the Standard notes it could be extended to as long as 22 months. The city has at least six data centers, two of them in Bayview-Hunters Point. Lurie said the city would use the time to build a framework around community needs, electricity rates, and clean water and air.
Who it affects: developers and colocation buyers planning new capacity inside city limits; the pause targets new development, not the data centers already operating. It also adds to a growing list of local fights over AI power demand, which is worth tracking if your company is choosing where to host GPU workloads.
Four quick answers before Monday
Can I still use Claude models in Replit?
Yes. Replit's CEO told TechCrunch that OpenAI is now the default provider for Replit Agent, but customers can still choose Anthropic's Claude models on the platform. The change mostly affects people who never touch the model setting, so open the picker and select Claude if you prefer its output for your projects.
Is Claude Fable 5.5 available in the API?
Not officially. As of October 18, 2026, Anthropic's newest documented Fable model is Fable 5.1. A Fable 5.5 identifier reportedly showed up in an Azure catalog screenshot and a LiteLLM model entry, and leakers expect a release soon, but Anthropic has published no model ID, price or date for it yet.
How much does DeepSeek's search API cost?
DeepSeek charges $0.75 per 1,000 successful Search requests, and the same price for Fetch. Empty results, failures and timeouts are free, and each tool is limited to 5 requests per second per account. Calls through DeepSeek Harness, Claude Code or Codex stay free until November 1, 2026.
How do I know if my OpenCode account was affected?
OpenCode says it has emailed every affected account and that users who received no email are not affected. Only about 1.4% of Console users, and only those who had linked outside providers, were exposed. Even so, rotating any keys you stored there is cheap insurance while the investigation continues.
Sources: TechCrunch, Replit and OpenAI (Oct 16–17) · TestingCatalog, Fable 5.5 leaks (Oct 17) · DeepSeek, Search & Fetch guide (Oct 2026) · DeepSeek, API pricing (Oct 2026) · MarkTechPost, V4.1-Flash-1017 (Oct 17) · Anthropic, employment report (Oct 16) · Reuters via The Star, Nova Scotia lease (Oct 16–17) · TokenPost, 95 MW phase (Oct 17) · Crypto Briefing, EverWind lease terms (Oct 17) · Notebookcheck, OpenCode breach (Oct 17) · TechCrunch, Meta terms (Oct 17) · Unite.AI, Meta effective date (Oct 17) · Motley Fool via Yahoo Finance, market close (Oct 16) · San Francisco Standard, moratorium (Oct 16)
