AI news, 18 October 2026: five stories with a date attached to them.

Five AI stories from the past week, and what links them is a calendar rather than a capability. Google's new frontier model is half price until 30 November. DeepSeek's web search API is free through coding tools until 31 October. OpenAI gave developers ninety days before four GPT-4 model IDs stop answering. Anthropic put six outside evaluators inside its own building, and signed an agreement with Singapore's digital ministry. Not one of these is a story about a model getting smarter. All five are dates you can put in a calendar this morning.

Every date and figure below comes from the primary source linked in its section. AcquihireTech writes this roundup for Indian service businesses doing ₹10L+ a month, so each item is filtered through one question: does it change what you do this month? The window is wider than usual, 13 to 17 October, and the reason is at the bottom along with the two stories that did not make it.

1. Google's new frontier model is half price until 30 November, and the second price is the real one

What happened. On 13 October Google released Gemini 4 Argon, which it calls its latest frontier model and aims at multi-step work in coding, finance, legal, cybersecurity and science. The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. Google states that once the introductory period ends on 30 November 2026, “the price of $4 per 1M input tokens and $20 per 1M output tokens will apply”. The model takes a 1M unified input-output token limit through the Gemini API. Google ranks it first on the Vals Index, reports an 87% score on the Embodied Reasoning (ERQA) benchmark, and says that on the Gray Swan indirect prompt injection benchmark attack success across 15 adaptive attempts fell to 0.7%. It is available globally through the API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Google Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Applications are open for 10,000 selected researchers to get six months free on Antigravity's highest usage tier, and cybersecurity defenders can apply to the Fairwind Program for access without cyber guardrails.

Why it matters. An introductory price with a published expiry is a budget trap wearing a discount's clothing. The number that governs anything you build this quarter is $4 and $20, not $2 and $10, because on 1 December your cost per call doubles and nothing about your product will have changed. Publishing the second price up front is more honest than most launches manage. It is still the price.

What it means for a service business. Cost your pilot at the 30 November number before you cost it at today's. If a workflow only clears its margin at $2 and $10, it does not clear its margin. Separately, the 0.7% prompt injection figure deserves more of your attention than the benchmark table above it: if a model is anywhere near a customer inbox or a CRM it can write to, indirect prompt injection is the failure mode that actually costs money. Read that number for what it is, which is Google's result on Google's chosen benchmark across 15 attempts, not a promise about your setup.

2. DeepSeek opened its web search API to everyone at $0.75 per 1,000 calls

What happened. On 17 October DeepSeek opened its Search and Fetch APIs to all users, describing them as “powered by our in-house search engine, serving hundreds of millions of requests daily worldwide”. The pricing is one line: “Each endpoint: $0.75 per 1,000 calls”, which is $0.00075 a call. Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are currently free for DeepSeek API users, with billing for those calls starting on 31 October 2026 at 16:00 UTC. The same release note says deepseek-flash is now served by DeepSeek-V4.1-Flash-1017, retires the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names, and publishes the weights on Hugging Face and ModelScope. DeepSeek describes the new version as having “stronger benchmark performance, faster task completion, and shorter outputs than the previous version”. Its benchmark scores appear on that page only inside images, so this roundup does not quote them.

Why it matters. Programmatic web search has been the expensive, rate-limited part of grounding a model in facts that are true today. At seventy-five cents for a thousand calls it stops being a line item worth a meeting and becomes plumbing. That is a bigger change to what a small team can ship than most model releases.

What it means for a service business. The practical use is narrow and good: a quoting workflow that checks a supplier's current listed price before it drafts the proposal, or a sales brief that reads the prospect's own site the morning of the call. The free window is the part to diarise. If someone on your team wires search into an agent through Claude Code or Codex this week, the meter starts at 21:30 IST on Friday 31 October, and usage that felt free in testing will not look free on the November invoice. Check what the thing costs per run before you point it at a thousand rows.

3. Anthropic put six outside evaluators inside its own building and gave up the right to veto what they publish

What happened. On 15 October Anthropic published an update on independent evaluators. Six employees of METR, a nonprofit that evaluates frontier models, have received company hardware, office space and access to internal tools and systems that Anthropic describes as roughly comparable to those of its own risk assessment teams. METR will publish its analysis of Anthropic's risk levels, incidents and practices. Anthropic states plainly that “Anthropic can't veto publication”, while keeping the right to require redaction within predefined categories such as customer data, privileged information and model architectures. The post sets out three tiers of access: narrowly scoped investigations with Anthropic-approved publication, publication without approval with some redaction rights retained, and the METR arrangement at the top. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August 2026 Risk Report. A broader partnership with Accenture and Faculty covers model evaluation, red-teaming, alignment assessments and safeguard testing, and Anthropic says early discussions are underway with others including Hugging Face. It compares the arrangement to bank supervision, where regulators place examiners inside the largest banks, and writes: “We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same.”

Why it matters. The 21 September edition of this roundup covered Anthropic's Accenture deal and called a billion dollars the reference price for credible outside scrutiny. A month later the money looks like the less interesting variable. The veto is the interesting one. An auditor who can be stopped from publishing is a consultant; an auditor who cannot is an auditor, and the redaction categories are where that distinction will actually be tested.

What it means for a service business. Nobody expects you to house an auditor. What travels downstream is the question, and it arrives in a procurement document: for each AI workflow that touches client data, who checked it, what did they check, and could they have said so if the answer had been bad? That is a one-page answer if you write it while the question is hypothetical, and a bad fortnight if you start writing it after a client asks.

4. Anthropic signed an agreement with Singapore's digital ministry, and published the usage numbers behind it

What happened. On 14 October Anthropic opened a Singapore office, its fifth in Asia-Pacific, and signed a Memorandum of Understanding with Singapore's Ministry of Digital Development and Information. Co-founder Ben Mann and Permanent Secretary Kai Fong Chng signed it at Singapore International Cyber Week. The MOU covers AI safety, security and model evaluation, research on AI adoption and its impact, and responsible adoption across the public and private sectors, with planned technical exchanges with Singapore's AI safety, security and assurance institutions. The post carries the numbers behind the move: Singapore ranks second of 121 countries in Claude.ai usage relative to its population, which Anthropic says is “nearly six times what its size alone would predict”; more than 1,500 engineers and technical staff at GovTech Singapore have adopted Claude Code; and Grab's BriX sales assistant saves its sales teams “approximately 40,000 hours every quarter”. DBS, GIC, Grab, Sea and Supabase are named as customers, and the Cyber Security Agency of Singapore is exploring how Claude could strengthen national cyber defence.

Why it matters. Keep the Grab figure. It is a vendor publishing a number about a customer, so read it as marketing, but it is specific, quarterly and countable about one named internal assistant, which is rarer in AI case studies than it has any right to be. Most of what circulates instead is a percentage with no denominator.

What it means for a service business. Two things. If you sell into Gulf or South-East Asian clients, model-evaluation language is now entering procurement from the government side rather than the vendor side, and that tends to arrive as a questionnaire. Second, the Grab number is a template for writing your own: hours saved per quarter, on one named workflow, measured the same way twice. A claim in that shape survives a sceptical buyer. A bare efficiency percentage with no denominator under it does not.

5. OpenAI gave four GPT-4 model IDs ninety days

What happened. On 14 October OpenAI notified developers that certain GPT-4 series models are deprecated and will be removed from the API on 14 January 2027, recorded on its deprecations page. The list is gpt-4o (snapshots gpt-4o-2024-08-06 and gpt-4o-2024-11-20), gpt-4o-mini (gpt-4o-mini-2024-07-18), gpt-4.1 (gpt-4.1-2025-04-14) and gpt-4.1-mini (gpt-4.1-mini-2025-04-14). The recommended replacement for all of them is gpt-6-luna. The page adds that “Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models.” Two nearer dates sit on the same page: twelve legacy GPT model snapshots shut down on 23 October 2026, and existing evals become read-only on 31 October 2026.

Why it matters. gpt-4o-mini has been the default cheap workhorse inside an enormous number of small production systems, including plenty built by agencies for clients who have long since stopped thinking about them. January feels far away in October. It will not in the first week of January, and a pinned model ID that stops answering does not degrade gracefully. It returns an error in whatever the model was quietly doing at 2am.

What it means for a service business. This is a half-hour job with a hard date on it, so do it badly now rather than well in December. Search your codebase, your automation platform, your Zapier and Make scenarios, and any notebook a contractor left behind, for those five model strings. Re-test the ones you find on gpt-6-luna and check the output format, because a classifier that returned a clean label can start returning a paragraph. The fine-tune line is the one that bites: if you have a fine-tuned gpt-4o-mini sorting enquiries, it dies with its base model on the same date, and replacing it means retraining rather than find-and-replace. That is a week of work hiding behind a one-line notice.

The stories we left out, and why this one is late

Two stories were cut. OpenAI announced an advertising format and measurement tools for ChatGPT, which is the most commercially relevant item of the month for anyone spending on performance marketing, and it is not above because openai.com returns HTTP 403 to every automated fetch from this pipeline. It did the same on 21 September and on 17 August. Reputable outlets report the details and are probably right; probably right is not the standard here. The deprecations story above survived only because developers.openai.com answers when openai.com does not. The second cut is Reflection AI's Beam, announced on 5 October with open weights promised later in the month: at the time of this run the weights release could not be confirmed at source, so it waits for the edition that can confirm it.

One more thing, said plainly because the dates give it away. This roundup last ran on 21 September. That 27-day gap was a broken publishing routine, not an editorial decision, which is why the stories above are drawn from 13 to 17 October rather than a strict 48 hours. The fix is in the repository this post was published from.

If you want the short version: cost anything new at Gemini's December price, diarise 31 October for DeepSeek search, and spend thirty minutes this week finding every gpt-4o-mini string you own. If the third of those turns up something you did not know was in production, that is the useful finding, and it is roughly where an AcquihireTech Systems Audit starts.