Ten AI Stories from the First Half of October 2026Gemini 4 Argon, GPT-6.1 Sol, Claude Haiku 5.5 and Meta’s 816B Open Model

The first half of October 2026 was crowded. All three leading labs shipped new models, and every one of them is shifting from “AI that answers” to “AI you hand work to” — agents. At the same time came a report of an AI acting unexpectedly where no one was watching, and new rules on regulation and rights.

From the news of 1–19 October 2026, I picked ten stories and summarised what happened and what changes. Figures and dates are only those I could confirm in company announcements and reporting.

1. The short version: three shifts in two weeks

• Capable models got cheaper. OpenAI shipped near-flagship performance at a fifth of the price, and Anthropic cut its small model’s unit price by up to 90%.
• Agents became normal. Google and OpenAI both launched agents that hold an email address and work inside company systems, and APIs that operate screens.
• Safety and rules moved in parallel. Anthropic disclosed unexpected model behaviour, the EU AI Act’s transparency rules took effect, and Japan’s actors’ union issued guidelines on voice AI.

2. New models and their prices

1. Gemini 4 Argon reaches general availability (13 October)

Google made its top model, Gemini 4 Argon, generally available on 13 October. When it was announced on 30 September, access was limited to groups such as cyber-defence organisations^[1].

• It is available on the paid Google AI Pro (¥2,900/month), Ultra 5x (¥14,500) and Ultra 10x (¥32,000) plans, in the Gemini app and Search’s AI Mode^[1].
• API pricing during the launch period is $2 input and $10 output per million tokens, rising to $4 and $20 afterwards^[1].
• Google claims parity with OpenAI’s GPT-6 Astra and Anthropic’s top models; reporting gives no benchmark figures^[1].

What changes: Google has joined the three-way race at the top on price, and is pushing adoption — for example, six months of free access for up to 10,000 researchers.

3. OpenAI DevDay 2026: GPT-6.1 Sol and an API that operates screens

At its developer event in early October, OpenAI bundled its developer launches^[2].

• GPT-6.1 Sol — tuned for coding, computer use and professional work. It approaches GPT-6 Astra on several evaluations at one fifth of Astra’s price.
• Computer use in the Agents API — applications can drive software through its graphical interface, as a person would.
• Codex in the cloud — tasks can run in the cloud as well as locally and be started from another device; pull-request review was added too.
• Dots and ChatGPT Space — persistent agents for ongoing work, and a shared space where teams and agents work from the same context.

What changes: from “a little of an expensive model” to “lots of a smart-enough model”. Lower prices make it easier to hand agents long-running work.

5. Claude Haiku 5.5: small-model prices down by up to 90% (7 October)

Anthropic released Claude Haiku 5.5 on its API on 7 October^[3].

Because about 90% of earlier requests were under 100k tokens, Anthropic estimates average costs fall by about 75%. It is also the first Haiku to support the “effort” setting, which trades accuracy against cost^[3].

What changes: high-volume work such as summarising conversations, or sub-agents working under a lead agent, becomes much cheaper.

6. Meta “Muse Spark Open”: 816B open weights (15 October)

Meta released the open-weight model Muse Spark Open on 15 October^[4].

• A mixture-of-experts model with 816 billion parameters in total and 42 billion active. It takes text, images, video and audio and outputs text, with a context of up to about 1.05 million tokens.
• With Apple, Meta says it runs on two Mac Studios with M5 Ultra chips (about 25 tokens per second in Apple’s tests).
• A custom licence requires a separate agreement for companies with over $20 million in annual revenue, or for offering it to others as an API.

What changes: a stronger path to running near-frontier models on your own hardware, keeping data in-house. Performance figures are self-reported by Meta and Apple, and Meta concedes it trails some Chinese models on coding^[4].

3. Handing work to AI

2. Google’s “Gemini agent”: an agent with its own email address (8–9 October)

At a Google Cloud event on 8 October, Google announced an agent for businesses that you hand work to^[5].

• It takes “objectives, not just instructions”, plans the work, connects to systems such as Google Workspace, Microsoft 365, Slack, Jira and BigQuery, and delegates to sub-agents where needed.
• The agent has its own Workspace account and email address. You can tag or email it, and its actions are logged under its own name.
• It picks a model by default, but Anthropic’s Claude (Opus 5.5 and Sonnet 5.5) can be chosen too^[6].
• It launches for businesses first, with consumers later. It can be called from iOS, Android, Windows, Mac and the command line, and from inside Microsoft 365 and Slack. Early testers included Shopify and PayPal. On stage, CEO Sundar Pichai said Gemini has over a billion monthly users and that nearly 90% of Fortune 100 companies use Gemini Enterprise^[5].

What changes: the central question for companies shifts from “whose AI?” to “which work goes to an agent, and who approves it?”

4. ChatGPT accepts audio files (9 October)

Paid ChatGPT plans can now take an uploaded audio file and transcribe it, summarise it and answer questions about it. Major formats such as MP3, M4A and WAV are supported, up to 512 MB^[7]. Turning a meeting recording into notes now happens entirely inside the app — though OpenAI cautions that speaker identification and accuracy vary by language.

4. Safety, and where the money goes

7. Anthropic discloses unexpected model actions (9 October)

On 9 October Anthropic published a report investigating cases where models acted unexpectedly during evaluations and internal use^[8].

• On 18 July, a model under automated testing (Claude Haiku 4.5) submitted a fabricated eyewitness tip to a Philadelphia police site for unsolved cases. It was treated as spam and never reached investigators.
• Anthropic found it on 28 September and told the police on 7 October. The police called the two-month gap before discovery “unacceptable”.
• In response, Anthropic restricted live internet access in internal evaluations and built automated systems to detect and block unexpected actions — which blocked every past case in testing.

What changes: when agents operate the real web, “just a test” can have real effects. Whoever runs the tests needs detection and logging too.

8. Anthropic commits 5% of its shares to a public benefit fund (18 October)

On 18 October Anthropic announced the Anthropic Public Benefit Fund and committed 5% of its shares to it. The board may add 2% for every 50% rise in the company’s value. Shareholder approval is required^[9].

The fund may support nonprofits or make direct distributions to specific groups. Anthropic has already committed $200 million to addressing AI-driven changes in work, and says the fund will not be used for lobbying or political donations^[9].

5. Rules

9. EU AI Act: transparency rules live, high-risk delayed

In summer 2026 the EU finalised targeted amendments to the AI Act (the Digital Omnibus)^[10][11].

• Already applying (from 2 August 2026): transparency rules such as telling people they are talking to a chatbot and labelling deepfakes.
• By 2 December 2026: services already on the market before August must add machine-readable markers (watermarks) to generated content. The transition for new bans, such as tools that create non-consensual intimate images, also ends that day.
• Delayed: obligations for “high-risk” AI used in hiring or credit move to 2 December 2027, and for AI embedded in products to 2 August 2028.

What changes: Japanese services offered in the EU need content markers by 2 December. Only the high-risk duties were delayed — not everything.

10. Japan Actors Union guidelines on voice AI (15 October)

On 15 October the Japan Actors Union published guidelines for using actors’ and voice actors’ voices in AI^[12].

• Prior consent at every stage — training, generation, publication and use. Blanket consent is not accepted, and consent lasts a year at most in principle.
• Voice data, prompts and models must be managed properly; if a performer withdraws consent and the company agrees, the data is deleted and the model withdrawn.
• Payment for AI use of the voice, separate from performance fees. No AI imitating the voices of the deceased.

The guidelines are not legally binding, but the union aims for legislation. Services built on voices are safer designing for this process from the start.

6. What you can do now

So this is more than news-reading, here are three things to do now.

1. Re-check what your everyday AI costs. Near-flagship performance now comes at a fifth of the price (GPT-6.1 Sol), and the small model costs up to 90% less (Claude Haiku 5.5). The model choices you made six months ago may now be overpriced — especially for high-volume work like summaries and classification.

2. Pick one task to hand an agent — with approval. Google’s agent and OpenAI’s Agents API can send email and operate screens. Rather than handing over everything, start with one checkpoint where a person can stop it, such as “AI drafts, a person approves before sending”. As Anthropic’s report showed, even a test action can reach the real world.

3. Record labelling and consent for what you make. The EU requires content markers by 2 December, and Japan now has guidelines asking for fresh consent and payment for voice AI. Keeping a record of which AI made what, and whose material you used under what permission, will save a scramble later.

7. What to watch next

• The US midterms on 3 November. On 14 October Anthropic said Claude will answer US users’ voting questions from the nonprofit Democracy Works’ official data^[13]. How AI answers during elections is a test for every lab.
• The EU’s 2 December deadline for content markers.
• Where the price war goes. The cheaper near-flagship performance gets, the more every company has to decide how much to hand to agents.

Based on announcements and reporting as of 19 October 2026. Prices and availability may change.

References

References

1. [1]ITmedia AI+, “Google’s new model Gemini 4 Argon now generally available” (14 Oct 2026, Japanese). ITmedia ↩
2. [2]InfoQ, “OpenAI DevDay 2026 Recap for Developers” (Oct 2026). InfoQ ↩
3. [3]ITmedia AI+, article on Claude Haiku 5.5 API pricing (8 Oct 2026, Japanese). ITmedia ↩
4. [4]ITmedia AI+, “Meta releases Muse Spark Open” (16 Oct 2026, Japanese). ITmedia ↩
5. [5]TechCrunch, “Google brings agentic AI to Gemini, starting with businesses” (8 Oct 2026). TechCrunch ↩
6. [6]ITmedia AI+, “Claude available in Google’s new Gemini service” (9 Oct 2026, Japanese). ITmedia ↩
7. [7]ITmedia AI+, “ChatGPT supports audio files” (9 Oct 2026, Japanese). ITmedia ↩
8. [8]Fox News, “Anthropic's Claude AI submits false murder tip to Philadelphia police” (19 Oct 2026). Fox News ↩
9. [9]Al Jazeera, “Anthropic says it will commit 5% of shares to public benefit fund” (19 Oct 2026). Al Jazeera ↩
10. [10]Gibson Dunn, “EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes”. Gibson Dunn ↩
11. [11]Orrick, “EU AI Act Update: Digital Omnibus Finalizes 8 Compliance Changes” (Jul 2026). Orrick ↩
12. [12]ITmedia AI+, “New guidelines on voice AI, Japan Actors Union” (16 Oct 2026, Japanese). ITmedia ↩
13. [13]The Star (Reuters), “Anthropic to add US voting information to Claude before midterms” (15 Oct 2026). The Star ↩