Issue2026.06

June 2026

Issue Zero: June 2026

Issue Zero of The Drift opens by walking backward. June 2025, then June 2026, laid side by side.

By 9 minute read

THE DRIFT

Issue Zero: June 2026

A monthly record of how AI is changing beneath your feet

OPENING NOTE

Issue Zero of The Drift opens by walking backward. June 2025, then June 2026, laid side by side.

A year ago, OpenAI, Anthropic, and Google all shipped frontier models inside the same ten-week window, each one wrapped in a different sales pitch. OpenAI called theirs reasoning models. Anthropic called theirs hybrid. Google called theirs long-context. Underneath the branding, all three companies were selling the same shift: a system that stops answering questions and starts finishing tasks.

A year later, that shift is no longer the pitch. It is the floor. This issue is built around a narrower question: what happens once reasoning, tool use, and long autonomy stop being features and start being infrastructure, who governs it, what gets deprecated, and which capabilities get cheap enough to disappear into ordinary software.

ONE YEAR AGO

On April 16, 2025, OpenAI put o3 and o4-mini into general use. The models arrived already wired to a browser, a Python interpreter, image and file tools, image generation, a canvas, automations, file search, and memory. OpenAI’s own language framed them as workers with an environment, not a better autocomplete.

Five weeks later, on May 22, Anthropic answered with Claude Opus 4 and Claude Sonnet 4. Sonnet 4 was priced at three dollars per million input tokens and fifteen per million output, the same rate as the Sonnet it replaced. Opus 4 cost five times as much: fifteen in, seventy-five out. Anthropic’s bet was specific. Make Sonnet cheap enough for daily coding and agent work, and let Opus be the model you reach for only when you need the ceiling.

By June, Google’s Gemini 2.5 Pro went generally available with a context window of one million tokens. Below 200,000 tokens it cost a dollar twenty-five in and ten out. Above that line, two fifty and fifteen. Google also collapsed a pricing distinction on Gemini 2.5 Flash, charging one rate whether the model was “thinking” or not. Long context, until then a novelty, became something a developer could budget for.

Underneath the three separate announcements sat one belief, unstated but shared: the next frontier would belong to whichever model reasoned best. That belief wasn’t wrong. It was just small. The larger and slower shift, visible only in hindsight, was that reasoning was turning into a pricing architecture. Developers had started grading models less on the quality of one answer and more on whether the model could survive being left alone inside a terminal, a codebase, or a multi-step task.

What none of the three companies were selling yet, because the category didn’t exist, was the assumption baked into every serious model release a year later: agent modes, code execution, browser and file access, memory, a visible deprecation calendar. In June 2025 those were features to announce. By June 2026, a model without them feels unfinished before it ships.

The anxiety was smaller too. In June 2025 the open questions were still close to the product: will the agent work, will it hallucinate mid-task, will enterprises trust it, will the open models catch up. A year later, the anxiety sits one layer up, in the release process itself: who approves a frontier deployment, how fast does the model underneath you get pulled, and whether the strongest capability shows up as public infrastructure or a restricted preview.

THE MODELS

June 2025’s three headline models, o3 and o4-mini, Opus 4 and Sonnet 4, Gemini 2.5 Pro, are one year old now. None of them is the news anymore. The news is that a tier below them, models built for volume rather than ceiling capability, has become the place where most production work actually runs.

What appeared

What arrived in the twelve months since was not a single model beating another. It was an operating assumption. A serious model in June 2026 is expected to reason across many steps, call tools, read and write files, work inside a codebase, and keep going without a person re-prompting it every turn. A year ago that was the pitch. Now it is the entry requirement.

What was deprecated or demoted

What got quietly retired were the deprecation notices OpenAI was already posting in mid-2025 for older realtime and audio preview models. At the time they read like routine housekeeping. Looked at now, they were the first sign that a model catalog is not a shelf. It is a dependency graph, and dependency graphs get pruned.

What moved from frontier to commodity

What moved from frontier to commodity is the June 2025 baseline itself. Summarizing, drafting, writing code, reading an image, following a multi-step instruction, the whole feature list from a year ago, is now the floor a model has to clear before anyone calls it serious. The word “frontier” migrated upward, toward reasoning depth, tool reliability, context length, price per useful token, and who is allowed to access the model at all.

What benchmark moved and what it meant

The benchmark that mattered a year ago was still mostly a leaderboard number. The benchmark that matters now is closer to a job review: can the model hold context, use a tool correctly, edit the right file, and finish the task without a human catching the failure downstream. Anthropic built its Claude 4 pitch around sustained performance on long-running work. OpenAI built o3 and o4-mini’s pitch around reasoning plus tools. Google built Gemini 2.5 Pro’s pitch around context and coding. All three were reaching for the same harder-to-fake measurement. They just didn’t have a shared name for it yet.

The model development that mattered most

The development that mattered most, visible only with a year of distance, was the rise of the model nobody put on a keynote slide: the cheap, fast, good-enough model. Sonnet 4 at three dollars and fifteen. o4-mini at a dollar ten and four forty. Gemini 2.5 Flash, repriced that same season to drop its thinking and non-thinking split. None of the three was the smartest model available. All three ended up doing the most work.

THE ECONOMICS

A year ago the price list looked like this: o3 at ten dollars in and forty out per million tokens, o4-mini at a dollar ten and four forty. Sonnet 4 held at three and fifteen. Opus 4 sat at fifteen and seventy-five, a five-times premium over its own sibling. Gemini 2.5 Pro split its pricing at the 200,000-token line, a dollar twenty-five and ten below it, two fifty and fifteen above.

Read today, that list is less a menu than a fossil record. Compression wasn’t an event that happened once in mid-2025. It became the operating condition. What cost fifteen dollars to run in April was competing with something that cost a tenth of that within months. Providers were not only racing on model quality. They were racing on which class of work they could make cheap enough to run by default, and which class they intended to keep metered.

That is still the split one year later. Deprecation calendars, cached-input discounts, fast tiers, premium tiers, and access restrictions now do as much work as the headline per-token price. A year ago the question was simple: what does this model cost? Now the better question is which kind of work a given provider is trying to make abundant, and which kind it is still trying to ration.

THE TOOLS

In 2025, Claude Code was a research preview. Nobody outside a waitlist had used it for long. OpenAI Codex had just come back, rebranded from an old code-completion tool into something closer to an agent. Cursor and Windsurf were teaching developers to expect a chat window that could see the whole repo, not just the file open in the tab.

The conversation among developers that year was still “can AI write code.” A year on, that question reads almost quaint. The live question is which workflow lets a model touch a codebase without leaving a mess behind it: which review loop, which permission model, which way of catching the change before it ships.

What died quietly in the meantime was the plain prompt box. A tool that offered nothing but a text field, no file access, no terminal, no browser, no memory, started to feel thin the moment repo-aware tools became normal. By 2026 the expectation is not a smarter chat window. It is a workspace: the repo, the terminal, the browser, the logs, and a model that can move through all of them.

The lesson a year of hindsight makes obvious: a weaker model wired into a good workflow beats a stronger model stuck in a blank chat box.

THE VIBE SHIFT

In 2025, the question people asked about a model was whether it was smart enough. A year later, the question is whether it can be trusted to act. That is not a small shift in tone. Awe is still there, but it now sits next to a second feeling: if a model can edit a file, call a tool, browse, and keep going for fifty steps unattended, a bad output stops being a wrong answer and starts being a wrong action.

There was a status change too. In 2024, someone using AI to code still had to explain themselves. By June 2025, the interesting people in the room had already stopped asking whether to use AI coding tools. A year further on, they are arguing about which agent, which context strategy, which review loop, and how much autonomy to hand over before nobody is checking the work at all.

That is the actual drift this publication is named for. Not whether the agents work. Who decides how fast the good ones get released, how quickly the older ones get pulled out from under you, and how much of the frontier an ordinary builder is even allowed to see.

WHAT TO WATCH

  • Whether restricted frontier previews stay a temporary exception or become the normal way a powerful model reaches the public.
  • Whether builders start treating model catalogs like a dependency they have to plan migrations around, the way OpenAI’s 2025 deprecation notices for realtime and audio previews now look in hindsight.
  • Whether cheap mid-tier models keep absorbing most production work while flagship models become the model you escalate to, not the one you start with.
  • Whether benchmarks keep moving away from static tests and toward long-running, tool-using, repo-scale tasks.
  • Whether the tools that win end up as standalone apps, IDEs, browsers, or something built into the operating system itself.

BY THE NUMBERS

Metric June 2025 baseline June 2026 meaning
Claude Sonnet 4 API price $3 input / $15 output per 1M tokens Strong coding and reasoning entered a durable mid-tier price band.
Claude Opus 4 API price $15 input / $75 output per 1M tokens The premium ceiling remained expensive, reserved for highest-value work.
OpenAI o3 API price About $10 input / $40 output per 1M tokens Deep reasoning was available but still metered as premium work.
OpenAI o4-mini API price About $1.10 input / $4.40 output per 1M tokens Low-cost reasoning became a serious production category.
Gemini 2.5 Pro context window 1M tokens Long-context analysis moved from novelty toward platform expectation.
Gemini 2.5 Pro API price $1.25/$10 below 200K tokens; $2.50/$15 above Long-context frontier work was priced aggressively against peers.
Gemini 2.5 Flash price change $0.30 input / $2.50 output per 1M tokens Google simplified thinking pricing and made fast reasoning easier to budget.
Claude Code public state Research preview/newly launched Terminal-native agentic coding became a serious category.

SOURCES