AI Signal – May 2026

TL;DR. The map matters more than the model. Seven things to carry into June:

• Pricing started moving at the edges in May; the flat-fee subscription assumption is no longer safe for the next twelve months.
• Google I/O made navigation, not capability, the small-operator problem.
• The harness layer matured across Codex and Claude Code; interaction emerged as its own model-architecture category.
• Compute moved into the Landscape as a strategic object, not background infrastructure.
• Token spend in the agent era is partly R&D; the discipline is capture, not minimisation.
• Multi-tool routing is now the architecture, not a hedge.
• Opus 4.8 (28 May) sharpened the model's judgment and added /goal and Dynamic Workflows; Opus 4.6 Fast Mode deprecated 29 June.

In late May, AI pricing started to shift at the edges. GitHub Copilot estimator screenshots circulating showed sharp increases for some workloads against the previous flat-fee equivalent.

The pattern is eye-catching. The honest reading is more nuanced. The shift hit the edges first: API users, out-of-harness consumption, enterprise customers. The in-harness subscription tier most small operators rely on (Claude Code, Codex CLI, Cursor) has not seen the same change yet, and a Claude Code Max subscriber today is still getting roughly the same value for money as six months ago. The direction of travel is set; the speed of travel depends on which side of the harness boundary a user sits. For a small operator the work is less about choosing the model and more about keeping a working map of which tools you use for what, at what cost, with which fallback. Three things actually mattered in May:

• AI pricing is shifting at the edges. Anthropic enterprise pricing, OpenAI's new Guaranteed Capacity commitments, GitHub Copilot, and Google's Ultra plan all moved towards usage-based billing through April and May. The change is real for API and enterprise users; in-harness subscription users have not seen the same shift yet. The direction of travel is set, and prudent planning means anticipating token budgets that move.
• Product surfaces sprawled out of any single map. Google's I/O confirmed the picture: a dozen new product names, an agent for your digital life with no release date, model tiers that nobody outside the company can keep straight. Navigation is now the unmet need.
• The case for access tiering got sharper. As compute stays zero-sum and security tightens, the small operator's likely future may be mediated access through product layers rather than clean APIs. The argument is structural, not yet a fait accompli.
• Opus 4.8 landed on 28 May, sharpening the model's agentic judgment and introducing explicit completion conditions (/goal) and parallel-subagent workflows. Arrived after publication; captured below.

That's the digest. The rest is the unpacking.

At a Glance

Key takeaway: the map matters more than the model. The model is good enough. The wrapper is mostly sold pre-built. What is left for a small operator to build is your own working map of the AI landscape: which tools you use for what, at what cost, with which fallback, inside which privacy boundary, and where you would never put AI at all. It is the part that compounds; it is the part that travels with you when the tools change. The four signals below all point at this from different angles.

Model Releases

---

The Landscape: what shipped

Models, Harnesses, Tools & Platforms — model and provider moves, compute capacity, interaction models, infrastructure, and the tools now available.

The most important release of the month was not a model. It was the realisation that frontier-lab economics had shifted underneath the conversation. Anthropic reported its first profitable quarter in late May, on a Q2 projection. It is the first time any frontier lab has posted a profitable quarter. There are caveats worth carrying: it is a projection, not a closed quarter; revenue accounting differs from OpenAI's; the result is partly profitable because the lab is supply-constrained and rationing access. None of that changes the headline. The bubble narrative that the labs would never serve agentic workloads at scale profitably lost ground in a single week. By the end of the month the picture had sharpened further: alongside the Opus 4.8 release on 28 May, Anthropic reported a $47 billion revenue run rate and closed a Series H round at a $965 billion valuation. The lab is now more valuable than OpenAI by that measure. The compute-as-kingmaker thesis moved from analyst commentary to a structural fact of the cost surface.

Google's I/O keynote dropped a dozen new product names into a single hour. Anti Gravity 2.0 (the agent system pulled out of the IDE and sold as the product), Spark (no release date), Omni (a Nano Banana moment for video, editing-first), Gemini 3.5 Flash (faster but materially more expensive than 2.0 Flash), and several more (AI Pro, AI Ultra, Gemini Business, Jules, Flow, NotebookLM). The list itself is the news. The sprawl is not a feature-launch problem to celebrate; it is a navigation problem to solve. A small operator now has to choose between a dozen overlapping Google AI surfaces with different pricing, different positioning, and several launched without release dates. Choosing what to use for what is now harder than using it. That gap is where the navigator role lives.

The single most consequential I/O change for a small-operator reader was buried inside the keynote: Google Search added a persistent-query axis. Agents inside Search can now hold a standing brief, monitor for new matches, and report back. The apartment search that used to be a one-off lookup becomes an open standing instruction. Most readers will encounter their first usable AI agent this way, inside a surface they already use, not by going looking for one. That changes the on-ramp for AI adoption: the agent finds you, you do not have to find the agent.

The cost-efficient middle is now competitive. Cursor's Composer 2.5 is the early example, benchmarking competitively with frontier coding models at materially lower cost, and similar moves are likely from other providers through the year. For any small practice doing coding-adjacent work (websites, automation scripts, data wrangling), the structural point is that routine coding tasks no longer need frontier-model pricing. Knowing which of your tasks actually need the top model, and which do not, is the discipline worth holding.

The harness layer kept maturing. OpenAI's Codex shipped a long-running objective loop (/goal), bringing it to parity with Claude Code on unattended multi-step work. The Codex team also pushed multi-surface (terminal, IDE, desktop, mobile, side panel) and an improved mono-thread context model that keeps a single workstream alive across sessions. Jason Liu on the Codex team published "Codex Maxing", a nine-pattern guide; the patterns line up almost exactly with the AgentOS Layers framework on our Practice framework page. On the Anthropic side, Claude Code has been moving in the same direction, with scheduled reflection, rubric-based grader agents, and multi-agent orchestration on shared infrastructure. The signal is not which feature you adopt; it is that the harness is now a stratified product surface, not a chat with tools.

Interaction emerged as a distinct landscape category. Thinking Machines Lab introduced an interaction-model architecture in May, trained from scratch for continuous, time-aware exchange rather than turn-based chat. The architecture pairs a foreground interaction model with a background model doing longer reasoning, browsing, and agentic work. It is early-stage, not yet a general platform choice, and worth treating as a category signal rather than a vendor recommendation. The wider point: interaction is becoming model architecture, not interface polish. If it generalises, the harness layer stratifies further: turn-based chat, persistent goal pursuit, and real-time interaction become three structurally different ways of working with AI.

Compute capacity became part of the landscape. Anthropic's compute partnership with SpaceX, OpenAI's Guaranteed Capacity programme, Google's cloud backlog moves, and Cursor's training of a new from-scratch model on Colossus 2 all pointed at the same thing: compute is now a strategic landscape object, not background infrastructure. For a small operator the effect shows up at the product surface: rate limits, usage-based billing, capacity-rationed features, and one-to-three-year commit deals all replace the old assumption that AI is a flat-fee subscription. Our Landscape framework page tracks AI Compute Infrastructure as a category in its own right for the first time this year.

One late-month addition: on 28 May, Anthropic released Opus 4.8. The headline is judgment: the model is four times less likely than 4.7 to let a significant flaw pass without flagging it, and more likely to push back on plans it judges to be unsound. Two new harness-layer additions landed with it: /goal (set a verifiable completion condition and step away; the model works to it without further prompting) and Dynamic Workflows in research preview (a fleet of parallel subagents coordinated within a single session, currently on Max and Enterprise plans). These are not a change to the thesis. They are evidence of it: the model gets better, the harness keeps maturing, and the small operator's job is still to map which of these capabilities are available to them now and which are direction signals to watch.

The Foundation: what is holding

Identity, Context & Connections (plus governance, readiness, trust) — the static AgentOS layers that decide whether the system can be used safely and repeatedly.

Four Foundation-level points carry from May. They are connected.

The first is the pricing shift visible at the edges through April and May. The moves were not synchronised, but the direction is convergent. Anthropic enterprise pricing went from a flat $200 per seat to $20 plus usage in April; the harness boundary now meters anything outside Claude Code and Claude Cowork, while in-harness subscription usage stays subsidised. GitHub Copilot moved to usage-based on 1 June; estimator screenshots circulating in late May showed sharp increases for some workloads against the previous flat-fee equivalent. OpenAI launched Guaranteed Capacity, a one-to-three-year compute commit deal that looks more like a cloud agreement than a SaaS subscription. Google's Ultra plan dropped headline price from $250 to $200 but added usage-based billing inside the agentic surfaces. None of this means the flat-fee subscription has ended for most small operators today; it does mean the flat-fee assumption is no longer a safe baseline for planning the next twelve months. Anticipate token budgets that move; build a stack that can absorb a single provider's pricing change without breaking your work.

The second, less-discussed point is what the subsidy end is the leading edge of. Three constraints are compounding: security restrictions on the most capable models, with the US Government moving towards more formal pre-release authority; compute as a genuinely zero-sum game, where the marginal cost of serving another user of a frontier model is high enough that the next iteration is unlikely to be materially cheaper than the last; and US Government strategic interest in who gets access to American-built tokens of intelligence. Together these point at a plausible equilibrium where the frontier model goes first to US national security, then to trusted defenders and US firms, then to KYC-cleared customers, and only later to everyone else, increasingly through pre-shaped product layers rather than clean APIs. The audience most affected by that shape is the Pandion reader: enthusiastic consumers, scrappy startups and small businesses, accessing frontier models through the chatbot and coding-agent interfaces of today rather than clean APIs. NLW on the AI Daily Brief named what had been building all month: the golden age of agent experimentation, which ran from roughly January to mid-2026, has come to a close. That is genuinely painful — experimentation is how practices discover the highest-value uses for agents, and non-technical operators especially need room to explore. It is also structurally healthier: pricing AI at its real cost is more sustainable than a subsidised bubble. For those paying attention, the slowdown period is an opportunity, not a retreat.

The third is in how to think about token spend. The convenient frame, "minimise consumption", is the wrong one in the agent era. Useful experimentation requires running things to see what happens. Every multi-step agent run, every alternative routing test, every prompt-tightening session burns tokens to produce knowledge that compounds across future work. The discipline is not to spend less; it is to capture what each spend taught you. One concept surfacing from practitioners in late May: 'agent debt' — the AI-era equivalent of technical debt. Hacking together an agent workflow during the experimentation phase and never cleaning it up produces conflicting system prompts, polluted memory, and overlapping tools. If you built agent workflows during the early-2026 experimentation rush, audit them now against what they actually cost and what they reliably produce. Anthropic shipped a /usage command inside Claude Code in May that breaks consumption down by skill, agent, MCP, or plugin (equivalent surfaces exist on OpenAI and Cursor). Run them once. The diagnostic side is useful. The harder discipline is the capture side: keeping notes on what worked and why, so a £200 token bill turns into £200 of saved time over the months that follow.

The fourth is in how AI governance is being formalised. Through 2025 governance was mostly internal policy: pick a risk framework (ISO 42001, NIST AI RMF, the EU AI Act if you serve European customers), document it, train the team. Through 2026 governance is starting to move into platform behaviour. The most capable models are now released with pre-deployment review, security-first access controls, and structural restrictions on who gets the strongest versions. A small operator does not have to track every policy move. The practical implication is simpler: assume that API key management, data classification at the prompt level, and an audit trail of what AI touched what data are no longer optional hardening; they are table stakes for any AI work involving protected client data. Pandion's Public/Private Wall pattern, set out on our Foundation framework page, is one shape of the practical response: separate client-confidential work onto privacy-first tools and keep the high-capability frontier tools on the public side.

One late-May addition sharpens the scarcity picture. Opus 4.8 introduced Fast Mode as an explicit pricing tier: 2.5x speed at double the standard cost ($10/$50 per million tokens vs the standard $5/$25). That is a deliberate design choice, not a temporary anomaly. Scarcity is now a dial you can turn: you can buy speed, but it costs tokens to do it. The Opus 4.6 Fast Mode is deprecated on 29 June; any workflow using /fast will default to Opus 4.8 pricing from that date. If you have cost expectations built around the old rate, check them before the end of June. Separately, Anthropic applied a +50% increase to weekly usage limits for Pro, Max, Team, and Enterprise plans through 13 July 2026. Worth noting not for the capacity itself but for what it confirms: limits are real, actively managed, and set to return to their previous levels. Plan your usage habits around the permanent level, not the temporary window.

The practical Foundation-level move that holds all four points together is multi-tool routing. For a Fortune 500 the conclusion is to negotiate compute commits directly. For a small operator the conclusion is more modest and more useful: do not depend on a single provider's subsidy floor, and do not depend on a single provider's mediated product layer remaining unchanged. A working setup that runs two or three tools side by side, routes each task to the right one, and treats tool diversification as resilience is now the prudent default, not an enthusiast preference. The deeper point is that the labs themselves now agree. Anthropic certifying 30,000 PwC professionals in Claude, and the broader lab-consulting pivot through April and May, are direct evidence: the labs have accepted that diffusion (not raw capability) is the binding constraint on AI's reach. Small operators doing what Accenture is being paid to teach Fortune 500 are on the right side of that curve.

A fifth signal arrived on 25 May, from outside the technology industry entirely. Pope Leo XIV published his first encyclical, Magnifica Humanitas, framing AI as this era's defining transformation and calling for it to be 'disarmed' — not abandoned, but freed from the competitive logics of military, economic and cognitive domination. For a small operator two practical points stand out. The labelling call is a governance direction likely to become a compliance expectation in professional and regulated sectors: AI-generated content should be clearly distinguishable from human-created work. And the concentration concern reinforces the multi-tool routing argument from a different direction: when capability is held by a handful of corporations, building a stack that does not depend on any single provider is sound practice whether the driver is pricing, strategic interest, or equity. 'The challenge,' Leo wrote, 'is not technological but anthropological.' The same argument, from a different altitude.

The Practice: how to work

Skills, Memory, Verification & Team Fluency — staging work, handoffs, rubrics, agent management, verification, and interaction as context transfer.

The most useful practitioner write-up of the month came from the Codex team itself: Jason Liu's "Codex Maxing" article, which set out nine working tips and bound them under a single integrating frame, don't break the loop. The vocabulary Liu reaches for (mono-thread, files-not-chat memory, voice as a way to brief richer context, side panel for parallel review, harness as a work system rather than a chat replacement) is the vocabulary Pandion has been using on our Practice framework page for several months. The patterns are not new; what is new is that the practitioners shipping the harnesses are now using the same words. The AgentOS layers framework is no longer Pandion-specific vocabulary.

The practical translation for a small operator is simpler than the literature suggests. Three habits do most of the work.

Stay on one thread per workstream. Treat the AI conversation as a long, evolving brief that compounds over weeks, not as a series of fresh chats. The good capture, the saved instruction, the standing pattern: all of it lives on the thread or in files the thread reads. The cost of a thread that gets long is small; the cost of restarting context every session is large. The Codex team's framing for this is "don't break the loop", which is exactly right.

Keep memory in files, not in chat. Anything you would need a future AI session to know goes into a markdown file you can paste from. A one-page brief about your business. A list of preferred tools and when you use which. A house style note. A short library of templates. The harness reads the files; the files survive the model. This is the part of AI fluency that compounds, and the part that travels with you when the tools change.

Brief like you are delegating, not pair-programming. Lead with the goal. State the constraints. Mark what is settled, what is provisional, and what is open for the agent to explore (the same handoff discipline a senior gives a junior). Define what "done" looks like explicitly, and tell the model what to verify before returning. Then let it run. Most disappointing AI output is a context failure, not a capability failure.

Three habits get you working. Beyond the habits, three disciplines mature the practice.

Agent management is now a learnable discipline. The three habits get you talking to an agent well. The next layer of discipline is what to do when the work is over: capture what you learned. A long multi-step agent run is half a deliverable and half an experiment. The experiment side produces knowledge: what prompts landed, what handoffs broke, where the agent went off-piste, what to do differently next time. The practitioner literature has started naming this directly: in the agent era, some token spend is research-and-development, not waste. The discipline is to capture each spend's learning as a structured note (in a file, of course) so it compounds into next month's better practice. Without this, every session reinvents what the last session learned.

Build a small library of skills. Skills are procedural knowledge made portable: a named recipe with prompts, files, examples, and verification rules that another session (or another team member) can load and run. Our Practice framework page treats this as load-bearing. It is the mechanism by which learning compounds beyond a single session and eventually scales to a team. A small operator's library does not need to be ambitious. A handful of named recipes ("how I do invoice reconciliation", "how I run a Friday review", "how I draft a client proposal") is more useful than a tagged-and-versioned system you never finish. Start by naming three things you do regularly and writing one file per thing.

Route work between tools deliberately. Loyalty to a single provider was an option when subsidy was abundant. In the new pricing environment, deliberate routing is a discipline. Match task to tool: frontier models for genuinely hard work, cost-efficient middle tiers for routine production, voice-friendly tools for high-context briefing, privacy-first tools for client-confidential work. The pattern is a multi-tool stack with consistent files-not-chat memory underneath, so routing does not cost you context. This is the operational form of the Foundation-tier point: routing is not a hedge, it is now the architecture.

The wiring is sold pre-built. The discipline of writing your work down so the wiring can act on it is yours. The three habits are how you start. The three disciplines are how you keep getting better.

One practitioner pattern from this month's captures is worth naming explicitly: the human sandwich. Dan Shipper at Every describes it as: you set the frame and success criteria; the agent drafts, searches, codes, and compares; you judge, extend, and decide what's next. This is the practical shape of human-agent collaboration in harnesses like Claude Code, Codex, and Cowork — neither turn-by-turn prompting nor fully autonomous delegation, but a repeating cycle where human judgment bookends the agent's execution. A related insight from the same source: shared team agents consistently outperform personal agents. An agent built around a recurring shared workflow — maintained once, used by everyone whose work intersects — produces better continuity, lower maintenance burden, and more durable team knowledge than a collection of individual replicas that each need separate upkeep.

A late-May addition to the practice layer is worth naming. The /goal command (Opus 4.8, available now across plans) takes the "brief like you are delegating" discipline one step further: set a verifiable completion condition at the start of a session rather than prompting step by step, and the model works toward it without further input. The habit worth building is learning to specify conditions precisely. What counts as finished? What constraints apply? What does the output need to contain? That specification work is the staging act, and it transfers directly into Dynamic Workflows (research preview on Max and Enterprise): a fleet of parallel subagents coordinated within a single session, each working a slice of a larger job simultaneously. If you are on a Pro plan today, /goal is available and Dynamic Workflows is a direction signal. The pattern is the same: define done, then delegate. Operators who are already practising condition-setting with /goal will find the step to fleet orchestration much smaller when it becomes available on their plan.

The Application: where it lands

Automations & Outcomes in your Domain — human premium, service-model change, new roles, continuous support, and outcomes in a real domain.

Where the human premium holds

The mainstream labour-market discourse caught up with something the practitioner conversation has been circling for two years. Alex Imas's essay What Will Be Scarce and Ezra Klein's response in the New York Times both named the same point: where the provenance of human creation or human service is part of the economic value, the relational sector grows in proportion to savings elsewhere. AI eats tasks; it does not automatically eat demand for human involvement.

For the Pandion reader this is not new advice, but the external articulation matters. It gives the small practice a defensible answer to the AI-displacement worry: the work most worth doing is the work where relationship, trust, accountability, translation, behaviour change, or provenance are part of the service. Continuous support that previously could not be afforded becomes viable. Personalised guidance that previously had to be templated can be made specific. Human escalation becomes a more important role, not a less important one. The same labour-saving that compresses some tasks expands the affordable surface of others. Most small practices already operate in this zone by default; the May discourse just gave it a name.

How services reorganise around agents

The most useful May reframe on this came from NLW's framing: agents make every job a startup. Two ideas in one phrase. The first is that work which used to be bounded by what a person could do in a week becomes bounded by what they can manage agents to do in a week. The cap moves outward. The second is that the new cap exposes work that was previously not affordable to do at all.

For a solo practitioner the second idea matters most. The things you previously could not justify spending an afternoon on (a deeper bit of client research, a custom note to a prospect, a properly-prepared meeting) become tractable. Backlog goes from finite-and-disappointing to unbounded. The ROI conversation that says "AI saves me three hours a week" misses the larger story. AI changes which work is worth attempting in the first place. A small practice with agent capacity reorganises around the new affordability: continuous client support that was previously impossible at solo scale becomes a feature; the personalised report that was previously a luxury becomes standard; the follow-up rhythm that was previously aspirational becomes routine. This is the deeper version of the Imas/Klein point: lower information cost does not just save existing tasks, it expands the affordable surface of work that has to be done well to be done at all.

The role families emerging

The other May thread worth carrying is the one about what new role shapes are appearing as agents handle more of the routine, mid-skilled, repeatable work. NLW spent several AIDB episodes through April and May testing role-family language: Agent Manager, Domain Operator, Context Engineer, Harness Engineer, Outcome Owner, AI Coach / Player-Coach, Navigator, Escalation Specialist. None of it is settled vocabulary. All of it is being used in the wild.

The Sequoia conversation with Jack Dorsey landed a simpler version: a small AI-native company has three roles, the builder/operator (a multi-functional individual contributor running the agents and the tools), the outcome owner (the human who carries accountability for what an agent or set of agents produces), and the player-coach (the senior who both does the work and teaches others how to do it; less management, more apprentice-style transfer). This maps cleanly onto the Pandion Capability framework's HR-equivalent layer, where the same shapes live as part of the team-design treatment on our Capability & Talent page.

The honest read is that nobody has these skills yet. Everyone is learning as they go. Universities are not yet shaped for it. Gen Z is entering a different graduate market than even five years ago. The small-business edge in this transition is agility: less hierarchy to unwind, faster experimentation cycles, the principal as the first navigator.

This thread is bigger than a single month's signal. A standalone Altitude piece in June, What Comes After the Org Chart, will take it on properly. For now, the line worth holding: the new question is not which jobs AI takes. It is which human roles become newly important when more work is routed through agents.

Framework Check

The four-tier framework (Landscape, Foundation, Practice, Application) held in May without strain. The four cross-cutting lenses (Scarcity, Staging, Interaction, Human Premium) are still the right ones. Scarcity sharpened materially this month with the pricing convergence, the access-tiering frame, and the structural shift of compute into the Landscape; the Opus 4.8 Fast Mode pricing and the Opus 4.6 deprecation are a late confirmation of the same direction. Human Premium was named by external voices we can cite. Staging continued to mature in the practitioner literature, and /goal is the most explicit formalisation of the staging pattern yet. Interaction earned its own beat in the Landscape, as a model-architecture category. Opus 4.8, arriving in the final days of the month, touched all four lenses at once: tighter judgment (Staging), explicit speed/cost dial (Scarcity), richer multimodal input (Interaction), and better reliability on the work where human accountability still matters most (Human Premium). The framework does not need a fifth tier.

Navigation is not a fifth tier. It is a cross-cutting capability that lives across all four. The Landscape gives you what to navigate. The Foundation gives you what the cost and access constraints are. The Practice gives you the habits. The Application gives you where it lands. The map is the small operator's edge.

What to do this week

---

The hardest sentence to land in a month like this is the closing one. The temptation is to reach for keynote-scale framing about Singularity foothills and golden ages of scientific discovery; every I/O and DevDay gives one for free. The more useful framing is the modest one: whatever golden age this technology might enable is our job to make true.

For a small operator, the way you make it true is more modest than the keynote framings allow for. It is the routing habit. It is the saved-context file. It is the small, durable harness. It is knowing who holds the map for your business and naming them. The next twelve months will be more uneven than the past three. Some surfaces will get more priced; some will stay subsidised. Product sprawl will accelerate. Access will continue to tier where security and capacity drive it. The work that compounds is not picking the right model. It is keeping the map current, whichever way any single provider moves.

The map matters more than the model. June's AI Signal will track the early experience of working through the new pricing reality. A standalone Altitude piece, What Comes After the Org Chart, will take on the role-families thread properly. Both are coming.

---

AI Signal is published monthly by Pandion Studio for anyone using AI as a core operating tool: solopreneurs, micro-organisations, small landscape and professional practices, and individuals using AI to organise their own life and admin. We read the AI firehose so you don't have to.

If you want help building the map for your own business, that's what AI Sessions are for.

FAQs