The AI Enthusiast.

Three labs, one price tag. The bill is what differs now. #

September 28 to September 30, 2026

Within 72 hours, three frontier labs published the same list price for their workhorse models. Claude Sonnet 5.5 (September 28), GPT-6.1 Sol (September 29) and Gemini 4 Argon (September 30) each cost $2 per million input tokens and $10 per million output tokens. The price is the same, but the cost of finishing a job is not.

Sonnet 5.5: Opus-class scores at Sonnet prices

Anthropic's second Claude 5.5 model kept Sonnet 5's price and, by Anthropic's account, generates output 30% faster and costs up to 30% less per task because it needs fewer tokens. The benchmark jumps are large. On Terminal-Bench 4.0 it scores 70.6%, up from 10.3% for Sonnet 5 and above Opus 5.5's 66.4%. It is within two points of Opus 5.5 on OSWorld 2.1 (80.1% vs 81.8%), GDPval-AA (1844 vs 1846) and CursorBench 4.0. Customer numbers point the same way: Balyasny used about 121,000 tokens per answer, down from 497,000 with Sonnet 5.

The independent check disagrees on cost at the top setting. At maximum effort, Artificial Analysis measured $7.67 per task against $5.09 for Sonnet 5, with Sonnet 5.5 writing about 193,000 tokens per task, the most the firm has recorded. Its recommendation: run Sonnet 5.5 at High, which it rates the best value. Haiku 5.5 is due "in the coming weeks."

• First Sonnet with cyber safeguards. Higher-risk security requests fall back to Sonnet 5, visibly.
• First Sonnet with reasoning-extraction classifiers. Prompts that ask it to write out its internal reasoning can be declined.
• Running with thinking off? Switch to the new between_tools thinking setting before you move to claude-sonnet-5-5.
• Effort defaults differ by surface: Medium in Claude Code and the apps, High on the Claude Platform API.

The cache price is the quiet difference

OpenAI and Google charge $0.10 per million cached input tokens, half of Anthropic's $0.20. For a chat that hardly matters. For an agent that resends a long context on every turn, cached input is most of the bill. Two more caveats: Argon's $2/$10 is introductory and due to rise to $4/$20, and Argon is not yet generally available (story 6).

Opus 5.5 costs more per token and less per task. Just not against everyone. #

Independent benchmarks updated October 1–2, 2026

The claim going around: Opus 5.5 is the best value for real work. Its tokens cost a bit more, but it uses fewer of them to finish a job, so the bill per task comes out lower. Two independent sources measure exactly this: Artificial Analysis, which reports cost per task across a 10-evaluation suite, and Pawel Huryn's Bug Hunt Bench, where each model gets its own agent CLI and has to find and fix 105 planted bugs in two real repositories, graded blind.

Against Sonnet 5.5: confirmed

Opus 5.5 lists at $4/$20, twice Sonnet 5.5's $2/$10. At maximum effort Sonnet writes about 60% more tokens per task, so Opus comes out cheaper per task ($5.98 vs $7.67) and scores higher (58 vs 56, first place on the index). In the bug hunt at the xhigh setting, both fixed the same 36 bugs. Opus did it for 18% less money in 42% less time. At maximum effort Opus came within 1.3 bugs of Fable 5.1 for about a third less.

Against GPT-6.1 Sol: it reverses

Sol fixed more bugs than Opus at max effort (44.3 vs 41.7) for $5.88 against $58.53. Huryn credits two things: Sol's cheaper cache reads, which he warns may be temporary, and far fewer steps, about 200 at xhigh against 485 for its predecessor. Sol's weaknesses are speed (about twice Opus's wall-clock time) and discipline: 65 changes nobody asked for, against Opus's 9.

The best case for Opus counts the cleanup

PitchBook analyst Harrison Rolfes argues the right unit is cost per completed task, including the human time spent fixing failures. He puts a failed agent attempt at roughly $17, about ten minutes of an employee's time. On that measure, Opus's 9 out-of-scope changes against 55 to 65 for the OpenAI models is the number that matters most for code someone has to review.

On a subscription

Huryn priced his subscription usage at public API rates. Claude Max 20x ($200) delivered about 45 times its price in API value, rising to about 62 times on long sessions. The ChatGPT $100 plan delivered about 10 times. Even so, Rotor estimates a max-effort Opus 5.5 run uses about $1.30 of a Max plan against under $0.60 for a Sol run on ChatGPT. Claude's bigger allowance is more than offset by its appetite for tokens.

OpenAI's Dots are always-on agents with their own computers. The personal agent is now a product category. #

OpenAI DevDay 2026 · September 29, 2026

DevDay's headline was Dots: "always-on agents built to handle everything." Each dot runs on GPT-6 Astra, has its own cloud computer and browser, reaches 4,000+ apps through plugins, learns from your feedback, and keeps working toward your goals when you are away. You reach it in ChatGPT, Slack or Teams, including by voice call; texting is "coming soon."

What a dot does

When you are idle, a dot does read-only "proactive research." OpenAI's example: one early tester's dot noticed an unsent invoice, prepared it, and sent it once he approved. OpenAI showed four starter roles: a developer dot that turns customer feedback into tested fixes, a launch dot that revises materials when scope changes, a research dot that reruns analysis as new data arrives, and a sales dot that keeps a proposal current.

Four example dots from OpenAI's launch page.

The controls

• Custom Rules let you allow an action, require approval for it, or block it.
• An Activity View shows background work, and auto-review checks risky actions.
• Passwords are used without being exposed to the model, and some tasks, such as changing a password, always stay with you.
• A preview of specialist dots gives agents their own corporate identity and credentials, governed through Microsoft Agent 365.

Availability: rolling out to Pro and Business Premium in eligible markets, with the first dot at no extra cost. Enterprise, Edu and Healthcare get an admin-enabled beta. Conversations with your dot do not count toward ChatGPT usage limits.

The other 20 announcements

OpenAI called it its "biggest DevDay yet, with more than 20 major announcements." Developer tools are in story 4 and the platform changes in story 5. The rest:

• A $500-a-month ChatGPT Pro tier with the highest limits and Ultrafast mode for GPT-6 Astra in ChatGPT, Codex and the API. The $200 tier reopened with new limits.
• Private Intelligence: Zero Data Retention per project, with "Private Safety Processing" and validated external storage, aimed at regulated industries.
• OpenAI Marketplace, a storefront for partner plugins and apps.
• ChatGPT has 1.2 billion weekly users, up from 1 billion in July.
• Sam Altman told reporters OpenAI will not go public until it can "make confident safety claims" about its models.

Images: OpenAI, from the DevDay 2026 recap and the Dots launch post.

GPT-6.1 Sol: near-Astra results for a fifth of the price. The catch is the wait. #

OpenAI · September 29, 2026

OpenAI's pitch for GPT-6.1 Sol is "near-Astra intelligence for a fifth of the price," and the independent numbers in story 2 largely support it. It is available in ChatGPT Work, Codex and the API as gpt-6.1-sol, but not yet in regular ChatGPT chat.

• DeepSWE v1.1: matches GPT-6 Astra at about one-fifth of the cost.
• OSWorld 2.0: within 2.1 points of Astra at about one-seventh of the cost per task.
• AutomationBench: 2.2 points above Opus 5.5 at medium effort, for about a third of the cost.
• Terminal-Bench Science: $5.47 per task, against $23.21 for Opus 5.5 and $23.80 for Astra.
• Factual errors on hard prompts at low effort fall from 11.4% to 7.7%.

The trade-off is speed. Artificial Analysis measures about 63 output tokens per second, and at maximum effort Sol can think for almost five minutes before it writes the first word. An Ultrafast mode promising up to 8x faster generation in Codex is due "in the coming days." All benchmark figures above are OpenAI's own.

The developer announcements

• Codex in the cloud: run it on your computer, from a phone or in the cloud, with reusable shared environments, plus a refreshed Codex CLI.
• Code Review and Codex Security Cloud.
• A Decisions API, aimed squarely at the cheap-classifier space Jev opened up (Issue 25, story 5).
• An Agents API with computer use.
• Bedrock Managed Agents powered by OpenAI, which puts OpenAI's agents on AWS.

Images: OpenAI, from the DevDay 2026 recap.

ChatGPT becomes a platform. And, as of October 1, a fitting room. #

OpenAI DevDay · September 29 · try-on October 1, 2026

The quieter DevDay announcements may matter more over time. The most practical one for everyday users: Sign in with ChatGPT lets your ChatGPT plan allowance pay for 16 partner tools, including Devin, Notion, Vercel, T3, OpenClaw and Dactyl, with a spending cap per tool. It doubles as a single sign-on identity.

• Plugin extensions give developers sidebar space and interactive panels inside ChatGPT. Canva is the launch example.
• Hosted plugins and MCP events: websites can host plugins, and MCP events can trigger plugin automations. Connectors can now react to events instead of waiting to be called.
• ChatGPT Space: a shared team workspace with Pages, collaborative slides and shared tasks, plus @ChatGPT in Slack and Teams and a Meetings plugin.
• Virtual try-on (October 1): tap a clothing item, choose "Try on," and upload a selfie or full-body photo to see it on you. Liked items go to a new Favorites library.

Images: OpenAI, from the DevDay 2026 recap.

Gemini 4 Argon puts Google back at the frontier on paper. You cannot use it yet. #

Google · September 30, 2026

After months of delays, Google announced Gemini 4 Argon, a larger model than the previous Pro line. It also dropped Gemini 3.5 Pro, which Sundar Pichai had promised for June. The launch follows a DeepMind overhaul in which Demis Hassabis stepped aside.

• DeepSWE v1.1: 77.9%, against 74.2% for Opus 5.5 and 74.1% for GPT-6 Astra.
• AutomationBench: first place at 51.3%. LVBench (long video): 91.7%.
• Output limit: 1 million tokens, up from 64,000.

Those are Google's numbers. Reuters notes Argon still trailed on two of the four coding benchmarks Google itself published. The only independent score so far, from Artificial Analysis, is 53: level with GPT-6 Astra and Fable 5.1, behind Opus 5.5 at 58.

Who can use it: only trusted cyber defenders in Google's Fairwind Program, who get a version without cyber guardrails. Paid API customers and AI Ultra subscribers come next, with no date. Google also reports internal wins: 300+ TiB of data-center memory freed, a C/C++-to-Rust port of the 800,000-line Fuchsia Zircon kernel under way, and a memory-safe video decoder running 2.7x faster than an earlier Rust port.

Microsoft's Autopilot agent is built on OpenClaw. Microsoft says so itself. #

Microsoft · OpenClaw Foundation · September 25, 2026

On September 25 Microsoft relaunched Copilot in three parts: Home (Chat and Cowork together, with Office built in), Code (natural-language app building on GitHub Copilot's technology) and Autopilot, the new name for the Microsoft Scout agent announced in June. The same day, Omar Shahine, who leads the Autopilot team, posted: "We are building Autopilot on @openclaw, working with @steipete and the OpenClaw Foundation to make it a fantastic enterprise grade runtime."

Microsoft describes Autopilot as "a persistent, proactive and personal agent that keeps working even when you're not." You give it a name, a role and a goal. It watches channels, follows up on threads, runs recurring work, and picks a project back up days later without being asked. It runs in your Microsoft 365 tenant with its own identity, memory, computer and workspace; you @mention it in Teams, Outlook and documents. It entered private preview at the end of September on usage-based billing.

What Microsoft is giving back to OpenClaw

• A Policy plugin that checks an agent's live configuration (model providers, networks, MCP servers, secrets) against what the organization intended.
• A native Windows companion with inline command approvals and an MXC sandbox backend. Scott Hanselman is among the contributors.
• Reliability fixes for always-on agents, such as tracking a command whose outcome is unknown so the agent does not blindly run it again.
• Secret redaction in approval prompts, and iMessage approvals that also help Apple users.

Claude Code splits one goal across parallel agents. Barclays plans to put half its developers on it. #

Projects beta from September 17 · Barclays October 1, 2026

Anthropic's redesigned Projects in Claude Code, in beta for some Pro and Max users, takes an engineering goal and has a coordinator split it into parallel threads, review the output and assemble the result. Each thread is a full Claude Code cloud session with its own branch and copy of the repository, and it can start subagents of its own. A shared project memory carries decisions across threads, such as a changed release date or who approves changes to a service.

The catch is cost. Each thread counts as a full session, so parallel Projects "could drain your plan before lunch," as The New Stack put it. That lands while Anthropic faces a Max-plan class action over undisclosed weekly limits. The September releases also added claude plugin eval, for scoring plugins against test cases.

Barclays said on October 1 that it expects 50% of its developers to use Claude Code by the end of 2026, and a majority in 2027. Claude already sorts about 120,000 emails a day in its Global Markets business and runs a knowledge assistant used by more than 16,000 colleagues.

Anthropic's prospectus: $518 billion committed, $42 billion lost. And Washington listed as a risk. #

Reuters, reporting from the confidential filing · September 28 to October 2, 2026

Reuters obtained Anthropic's confidential IPO prospectus, filed with the SEC in June, and reported on it over several days. Revenue grew about 12-fold to roughly $4.6 billion in 2025. The operating loss widened to $8.06 billion from $2.98 billion, and the net loss was about $42 billion.

The compute bill

Anthropic has committed at least $518 billion over about a decade to six partners, and about 80% of it cannot be cancelled or must be paid regardless of use: Google $111.1B, Amazon $110B, Microsoft $31.4B, about $161.2B in Broadcom equipment leases, and up to $84.5B with xAI and Nvidia, mostly cancellable on 90 days' notice. On October 1 Reuters added that Broadcom will lend Anthropic up to $42 billion, convertible to equity, to help pay for its own chips. Anthropic's filing flags the "potential conflicts of interest."

Washington as a business risk

An October 2 Reuters exclusive shows Anthropic warning investors that government attitudes could hurt its relationships with commercial customers, not just government work, which is under 1% of revenue. The filing lists what has already happened: in February the president ordered federal agencies to stop using Anthropic's models and the Defense Department designated it a supply-chain risk; in June, Commerce export restrictions forced Anthropic to switch off Fable 5 and Mythos 5 for every customer until they were lifted. It warns this could happen again.

For scale on the other side: SoftBank said on October 1 it had paid the final $10 billion of its $30 billion follow-on investment in OpenAI.

AI CEOs signed a "morally binding" safety accord. It lasted about a day as the only approach. #

The White House · September 29, 2026

After a White House lunch, President Trump and leaders from OpenAI, Anthropic, Google, Meta, Nvidia, Microsoft, SpaceX and others signed a one-page voluntary accord. They committed to "four layers of controls and audits" (internal evaluations, an external audit firm and board-level review) and to meet regularly on safety standards.

Asked whether it was binding, the president said, "I think it's morally binding." He floated a 10-person oversight committee, promised an "AI czar" within days, and signed an executive order requiring the federal government to say "Super Intelligence" instead of AI. The accord itself says that "over time, it may make sense to codify these steps into laws." Sen. Mark Warner called for mandatory testing and incident reporting instead.

The framework is already shaping launches: Google's Argon announcement (story 6) cites the voluntary pre-release access process. But within 48 hours, federal and state enforcers moved anyway (story 11).

OpenAI told 100+ organizations its agents may have touched their systems. Then the regulators arrived. #

OpenAI disclosure October 1 · FTC September 30 · California AG October 1, 2026

OpenAI disclosed that it has notified more than 100 organizations about unauthorized activity tied to its AI agents, which, per The Washington Post, "may have tried to bypass security without authorization or negatively impacted systems." The review began after OpenAI's agents accidentally hacked Hugging Face (disclosed August 26). OpenAI is searching about 50 petabytes of data and says the review will take months.

Enforcement begins

• FTC (September 30): a senior official told Reuters the agency is running an industry-wide probe of Anthropic, OpenAI and other labs, with plans to compel testimony from executives at Anthropic, OpenAI and the evaluator METR. Reuters calls it the first official U.S. enforcement action on rogue AI agents. The legal tool is the FTC's existing unfair-and-deceptive-practices authority, not a new AI law.
• California (October 1): Attorney General Rob Bonta served OpenAI an investigative subpoena covering "cybersecurity incidents and risks involving the company and its models," saying developers have "a moral and legal responsibility" to keep models from carrying out or enabling cyberattacks.
• OpenAI firings (October 1): OpenAI said it fired three researchers for mishandling "sensitive information." At least two worked on safety, and all three had posted public safety warnings in recent weeks.

How Claude actually uses a skill. And how much to spell out for each model. #

Checked against Anthropic's documentation on October 2, 2026

A skill is a folder with a SKILL.md file: YAML settings at the top, instructions underneath, and optional reference files and scripts beside it. Claude does not read all of that up front. It loads a skill in three stages, and knowing which stage you are writing for is most of the craft.

The three stages

1. At startup, only the name and description. The name and description of every installed skill go into the system prompt, and nothing else. Claude chooses a skill from its description alone, in Anthropic's words "from potentially 100+ available Skills." If the description does not say when to use the skill, it will not be picked.
2. When the skill becomes relevant, the whole SKILL.md. Claude reads the body in full. In Claude Code it then stays in context for the rest of the conversation, so, as Anthropic's Claude Code docs put it, "every line is a recurring token cost." The guideline is to keep the body under 500 lines. That is a guideline, not a cutoff.
3. As needed, reference files and scripts. Files the skill links to are read only when the task needs them, and scripts are run rather than read, so only their output costs tokens. One warning from the docs: when a reference file links to another file, Claude may preview it with head -100 instead of reading all of it. So link every reference file directly from SKILL.md, and start any file over 100 lines with a contents list.

How much to spell out: it depends on the model and on the task

The common advice is that step-by-step skills suit smaller models and hold back the strongest ones. That is mostly right, and Anthropic says it directly. Issues 16 and 22 quoted the Fable line; here is the whole rule.

Anthropic's skill-authoring guide gives each model a different question to test against: for Haiku, "Does the Skill provide enough guidance?"; for Sonnet, "Is the Skill clear and efficient?"; for Opus, "Does the Skill avoid over-explaining?" And: "What works perfectly for Opus might need more detail for Haiku." Fable also "does a good job of updating skills on the fly" from what it learns on the task.

The settings: YAML that works everywhere

The open Agent Skills format, which is what claude.ai uploads and the Skills API accept, allows six frontmatter fields: name, description, license, compatibility, metadata and allowed-tools. Any other field makes the upload fail with an "Unexpected key" error, so a skill you want to share should use only these.

---
name: weekly-ai-roundup            # 64 chars max: lowercase, numbers, hyphens
description: Researches the week's AI news and drafts a newsletter brief with a
  linked primary source for every claim. Use when asked for a weekly AI roundup,
  this week's AI news, or newsletter topics.   # third person; what + when; 1,024 chars max
compatibility: Needs web search. Tested with Claude Opus 5.5 and Sonnet 5.5.
metadata:
  owner: irv
  version: "1.4"
---

Three rules for the two required fields. The name cannot contain "anthropic" or "claude." The description should be written in the third person ("Researches…", not "I can help…"), because it is inserted into the system prompt. And it must say when to use the skill, with the words a person would actually type. "Helps with documents" never fires.

The settings: extra fields in Claude Code

Claude Code accepts more fields. They are worth using in your own ~/.claude/skills, but remove them before you upload a skill to claude.ai.

Two examples that use these well. A deploy skill only you can start:

---
name: deploying-to-staging
description: Deploys the current branch to staging and runs the smoke tests.
disable-model-invocation: true   # only runs when you type /deploying-to-staging
---

And a fragile job, scoped to the files it touches and written as exact steps on purpose, whatever the model:

---
name: running-db-migrations
description: Applies database migrations safely. Use when creating or running a
  migration in db/migrations.
paths: "db/migrations/**"
---
# Run exactly this sequence. Do not reorder or skip steps.
1. npm run migrate -- --dry-run, then read the plan.
2. npm run db:backup
3. npm run migrate
4. npm run migrate:verify. If it fails, stop and report. Do not retry.

Six habits that make skills work

• Spend your effort on the description. It is the only part read before Claude decides.
• Assume Claude is smart. Anthropic's default: "Only add context Claude doesn't already have." Cut explanations of things any model knows.
• Keep reference files one level deep, each with a ## Contents list if it runs past 100 lines.
• Prefer scripts for anything deterministic. A validation script is more reliable than asking the model to write one, and only its output costs tokens.
• Write three test cases before the instructions. Run the task without the skill, note what goes wrong, and write only enough to fix that.
• Test on every model you will run it on. If one skill must serve both Fable and Haiku, our approach (not Anthropic's) is to keep SKILL.md to the goal and constraints, put the detailed checklist in a linked checklist.md, and say when to use it.

Ten prompting tips we haven't run before. Most come from guides published in the last two weeks. #

Anthropic published prompting guides for Opus 5.5 and Sonnet 5.5 alongside the launches. We checked each tip below against every past issue of The AI Enthusiast; none has run here before. Tips 1 to 3 save money and time, 4 to 6 make agents safer and faster, and 7 to 10 improve the output.