Anthropic and OpenAI both dropped token prices on 22 September. The cost per finished task barely moved.
Anthropic released Claude Opus 5.5 on 22 September at $4 per million input tokens and $20 output, down from $5 and $25. The more consequential change is cache reads, which fall from $0.50 to $0.20 per million — a 95% discount against uncached input, up from 90%. OpenAI shipped GPT-6 Sol and GPT-6 Luna the same day with headline cuts of roughly half: Sol to $2 in and $10 out, Luna to $0.10 and $0.50, alongside a prompt cache that keeps the same 90% read discount but adds a 30-minute reuse window and explicit cache breakpoints.
The independent numbers say the money did not move as far as the price list. Artificial Analysis measured Opus 5.5 at max effort at 58 on its Intelligence Index v4.3.2 — the highest it has recorded, leading six of ten constituent evaluations — but the model spends about 119,000 output tokens per index task against roughly 73,000 for Opus 5, so cost per task comes out level despite the 20% cut. For GPT-6 Sol and Luna the cost drop is real — Sol from $1.99 to $1.06 per index task, Luna from $0.18 to $0.07 — but AA puts both level with GPT-5.6 on intelligence, with Luna's Coding Agent Index falling two points and GDPval-AA Elo down about 100 for Sol. One further caveat for anyone modelling this: OpenAI's comparison is reported to be drawn against GPT-5.6 promotional pricing rather than a published standard rate, a point the company itself made when it said the new rates carry no promotional expiry. Price per token is now a poor proxy for price per job, and only a per-task measurement tells you which way a switch actually goes.
Your Jira MCP server may have been answering as you since July
GitHub's advisory database published around 25 advisories against mcp-atlassian on 22 September, the widely deployed Jira and Confluence MCP server. The critical one is GHSA-wrhw-j3f9-8vc6, CVE-2026-77244, scored 10.0 on CVSS v3.1 — the maximum, on an unauthenticated network vector with changed scope: the token verifier returned a valid access token for any non-empty string, the OAuth proxy is off by default, and a request carrying no Authorization header at all is not rejected — the server falls back to environment configuration and answers using the operator's own Jira API token. Most of the remaining advisories are one root cause: upload_attachment and download_attachment accepted an unvalidated file path, allowing arbitrary reads of server-local files and exfiltration into a Jira ticket.
Exposure needs the server run over HTTP and network-reachable, which is the documented quickstart for a shared deployment; stdio-only local use is not affected. The uncomfortable part is the timing. Version 0.22.0, which fixes all of it, shipped on 10 July. Anyone pinned below it has been exposed for ten weeks and only received a dependency alert this week. If your agent platform mounts third-party MCP servers, the advisory feed is not an early-warning system — a fix landing in a changelog weeks before its CVE is now the normal shape of this. Five LightRAG advisories went out in the same ten minutes with the same timing problem: a different bug class, but a fix released on 31 July.
47 of Britain's 48 police forces run through a US hyperscaler
Computer Weekly analysed 3,822 public DNS records, resolved to their owners through RDAP, across every UK police force, and found that 47 of 48 have at least one hyperscaler connection, and 46 route email and identity through Microsoft 365. Microsoft accounts for 773 routable records against Amazon's 161 and Google's two. Only the Ministry of Defence Police runs with no hyperscaler at all. The outlet excluded TXT records, so its figures are a floor rather than a ceiling.
This is an original count rather than a vendor survey, which is what makes it worth reading twice. Every one of those email and identity paths is a data flow that a US disclosure demand can in principle reach, and the concentration is not a procurement decision anyone appears to have taken deliberately — it is what standardising on the obvious option produces over a decade. The same question sits under most European public-sector estates, and almost nowhere has it been measured. Knowing which of your own records resolve where is cheap; discovering it during an incident is not.
Quick Hits
- A Finnish AI cloud raised $189m and declared itself a unicorn. — Verda, formerly DataCrunch, closed an oversubscribed Series B led by Emergence Capital on 22 September, taking total funding past $450m across equity and debt. It states a $165m annualised run rate for July 2026 and more than 250MW of operations planned for 2027 — its own unaudited figures, and a run rate is not revenue. The valuation above $1bn is likewise the company's own framing rather than an independent mark.
- The EIB put €20m into a Dutch lithography firm, and the coverage rounded the round. — The bank's own release on 22 September states a €20m convertible loan to Morphotonics as part of a Series B “of over €40 million” also involving BOM and Invest-NL under InvestEU. Several outlets reported a flat €40m. The primary gives a floor, and half the disclosed total is debt rather than equity.
- Austria proposes a full environmental assessment for data centres above 100MW. — The government sent a draft revision of the Umweltverträglichkeitsprüfungsgesetz into consultation on Monday, running to 1 November, under which data centres of 100MW and above — or 50MW in protected areas — would require an environmental impact assessment, heise reports. The same draft streamlines approval procedures generally and raises new hurdles for environmental organisations and citizens' initiatives. It is a consultation draft, not law.
- The UK's AI Security Institute started publishing its evaluation transcripts. — AISI is releasing results through EvalEval's Evaluation Cards using the Every Eval Ever schema, with configuration and transcript-level detail, announced 22 September. The first drop covers five benchmarks across six frontier models, all a generation old. It is also usable infrastructure for reconciling the vendor-versus-independent benchmark gaps in today's lead.
- Open WebUI now accepts tokens you have revoked, if Redis is down. — v0.11.4, released 21 September under an explicit security advisory, makes the revocation list fail open — sign-out may not take effect during a Redis outage — and turns cookie forwarding into an opt-in switch, so an instance relying on it must re-enable it after upgrading. Live sessions are now torn down on revocation, which they previously were not.