The agent era, dated and sourced.

Matthew Butterick and the Joseph Saveri firm sue over training on public repos without honoring open-source licenses — the first major lawsuit of the generative-AI era. By July 2024 a judge trims 22 claims down to two, but the case defines the legal risk conversation every engineering org has about AI code.

GitHub Copilot

After eight years and 500K monthly developers, Kite folds and open-sources its code; founder Adam Smith concludes the ML of its day 'was not good enough' and developers would not pay. Nine days before ChatGPT, it marks the exact boundary between the old bet and the new one.

Four MIT students at Anysphere ship an editor where the AI is the interface, not a plugin, in March 2023; an $8M seed led by the OpenAI Startup Fund follows in October. It looks like a niche fork — within two years it is the fastest-growing dev tool in history.

Cursor

Multimodal, top-10% bar exam, and a napkin-sketch-to-website demo: GPT-4 makes GPT-3.5-era code generation look like a toy. It immediately becomes the engine of Copilot X, early Cursor, and most serious coding workflows of 2023.

The model that started the category is deprecated with a March 23 shutoff — 'our latest models are now our best models for coding' — pushing developers to GPT-3.5-Turbo and GPT-4. Specialized code models lose to generalists, for a while; the Codex name returns as an agent in 2025.

OpenAI Codex

A week after GPT-4, GitHub shows Copilot expanding from completions into chat, pull-request descriptions, docs Q&A, and a CLI helper. Most of it ships over the following year, and 'chat in the editor' becomes table stakes for every competitor.

GitHub Copilot

Hugging Face and ServiceNow release StarCoder under an OpenRAIL license — 15B parameters, a trillion tokens of permissively licensed code, 80+ languages, with the training data documented. It sets the transparency bar the open code-model wave builds on.

Rolling out to Plus subscribers, Code Interpreter lets ChatGPT write and execute Python against uploaded files — the first mainstream taste of a model that runs its own code and reads its own errors. The write-run-fix loop later becomes the heart of every coding agent.

Anthropic's first broadly public model reads hundreds of pages per prompt and scores 71.2% on HumanEval, up from 56%. The 100K window makes whole-codebase questions practical and starts the long-context race that defines coding-model usefulness.

7B, 13B, and 34B code models under the Llama 2 community license put capable local code generation on a single GPU. Self-hosted coding assistants stop being a research hobby and become a real deployment option.

Princeton's benchmark uses 2,294 real issues from real repos; the best model of the day (Claude 2) resolves 1.96%. That number becomes the era's most-watched scoreboard — from under 2% to over 80% in roughly two years.

Google's answer to GPT-4 ships natively multimodal in three sizes, with Ultra claiming the first above-human MMLU score. Google is now fully in the frontier race that its coding tools — Gemini CLI, Jules, Antigravity — will draw on.

Chat lands for every paid Copilot subscription in VS Code, Visual Studio, and JetBrains — free for students and OSS maintainers. Conversational coding stops being a preview feature and becomes part of the default developer toolchain.

GitHub Copilot

Google demonstrates production long-context at 1M tokens — 30,000+ lines of code in one prompt — via MoE architecture. 'Fit the whole repo in context' changes from fantasy to roadmap item for every coding tool.

The top Copilot tier adds codebase-aware chat on github.com, docs indexing, and PR summaries at roughly 4x the individual price. The market learns what enterprises will pay for AI coding when it knows their code.

GitHub Copilot

Opus, Sonnet, and Haiku establish Anthropic's three-tier naming and put Claude ahead of GPT-4 on most public evals for the first time. The Sonnet tier's price-to-coding-quality ratio quietly becomes the foundation of the coming agent wave.

AWS consolidates its developer AI under the Q brand with agents that can implement features and modernize Java apps. Amazon's real bet on agentic coding — Kiro — is still a year out, but the CodeWhisperer chapter closes here.

At $3/$15 per Mtok it beats Opus on coding (64% vs 38% on Anthropic's internal agentic eval) and ships with Artifacts for live code preview. Bolt, Cursor, Windsurf, and Copilot's multi-model tier all standardize on it — the first time one model visibly powers a whole product category.

a16z and Thrive lead the round as Cursor's word-of-mouth growth turns a VS Code fork into a breakout. The valuation looks rich for an editor; it will look quaint within a year.

Cursor

93 developers hand-screen the benchmark down to 500 solvable, well-specified tasks after finding the original systematically underestimated models. 'SWE-bench Verified' becomes the number every frontier launch quotes — with all the over-optimization risk that implies.

Launched in September 2024 on Claude, Replit's agent takes non-programmers from description to deployed application in minutes. It rides the coming vibe-coding surge from $2.8M to $150M+ annualized revenue within a year — proof the agent market is much bigger than professional developers.

Trained to think before answering, o1 hits the 89th percentile on Codeforces and shows that inference-time compute is a new scaling axis. Reasoning becomes the ingredient that makes long agentic coding runs actually converge.

StackBlitz pairs Claude 3.5 Sonnet with WebContainers so full-stack apps generate, run, and deploy entirely in a browser tab — launched with a single tweet. Its growth curve announces the vibe-coding consumer market louder than any keynote.

Anthropic jumps SWE-bench Verified from 33% to 49% and lets the model drive a screen — cursor, clicks, typing — at the same price. Computer use is rough (14.9% on OSWorld) but establishes the agent-does-what-you-do direction every vendor follows.

At Universe, GitHub adds Claude 3.5 Sonnet, Gemini 1.5 Pro, and o1 to Copilot, ending the OpenAI-exclusive era of the biggest AI dev tool. Model choice becomes a checkbox feature everywhere, and the model layer starts commoditizing beneath the tools.

GitHub Copilot

Alibaba's code model tops the open-source boards (EvalPlus, LiveCodeBench, Aider 73.7) at a size hobbyists can self-host. It is the first loud signal that Chinese open-weight labs will contest the coding frontier, not trail it.

Qwen 3.6

Windsurf launches as a standalone agentic IDE whose Cascade agent tracks project-wide context, edits across files, and runs terminal commands in 'flows'. It positions Codeium as Cursor's most direct rival — and sets up 2025's wildest acquisition saga.

Windsurf

MCP standardizes how models connect to tools and data: servers expose capabilities, any client can consume them. It launches quietly with a handful of reference servers; within a year it is the de facto integration layer of the entire agent ecosystem.

Cognition prices Devin like a contractor, not a tool: unlimited seats, Slack-native, aimed at teams delegating first-draft PRs and backlog fixes. The price point tests what autonomous work is worth — and lasts four months.

Windsurf

2,000 completions and 50 chat requests a month, no card required, with GPT-4o and Claude 3.5 Sonnet included. The reference product going freemium resets the floor price of AI coding to zero and pushes rivals toward premium agent features.

GitHub Copilot

The 671B MoE (37B active) matches GPT-4-class performance with a training bill two orders of magnitude below Western assumptions. The efficiency claim rewires the economics debate a month before R1 makes it a market event.

DeepSeek V4

R1 matches OpenAI's o1 on key benchmarks with published weights and API pricing of $0.55/$2.19 per Mtok against o1's $15/$60. Every operator's cost model — and every lab's moat assumption — gets rewritten in a week.

DeepSeek V4

R1 panic produces the largest single-day market-cap loss in stock-market history as investors reprice AI compute demand. For practitioners the lesson lands differently: frontier-adjacent coding capability is now nearly free, and the Chinese open-weight wave is structural.

DeepSeek V4

A throwaway tweet — 'fully give in to the vibes, embrace exponentials, and forget that the code even exists' — gives the practice a name, 4.5M views, and eventually a dictionary entry. Naming the workflow legitimizes a market segment the tools were already serving.

GitHub's 'agent awakens' release lets Copilot iterate on its own output — run commands, watch tests, self-correct in a loop — reaching VS Code Stable with MCP support that April. The largest install base in the category shifts from assistant to agent.

GitHub Copilot

Anthropic ships a terminal-native agent that searches, edits, tests, and commits, alongside the first hybrid reasoning model. It looks spartan next to the IDE products; the terminal turns out to be exactly where serious agent work wants to live.

Claude Code

A thinking model with a 1M-token window that tops WebDev Arena and the app-building evals, though Claude 3.7 keeps the SWE-bench edge. After a year of also-ran coding models, Google is competitive again — three months before Gemini CLI.

Sam Altman announces MCP support across the Agents SDK, ChatGPT desktop, and the Responses API — 'people love MCP.' A competitor adopting Anthropic's standard settles the agent-integration format war barely four months after it started; Google follows within weeks.

Cognition moves to pay-as-you-go ACUs and claims 83% more completed junior tasks per compute unit. A 96% cheaper front door in four months tells the market what the first year of autonomous-agent pricing was worth.

Windsurf

Codex CLI ships on GitHub alongside o3 and o4-mini — a terminal agent that edits files and runs commands locally. Two months after Claude Code's preview, the terminal-agent format has both frontier labs' names on it.

OpenAI Codex

The reported deal — OpenAI's largest acquisition to date — values the Codeium/Windsurf business at 2.4x its private mark eight months earlier. It also plants the question that detonates in July: what happens to a coding startup whose main model vendor is a rival?

Windsurf

Each task gets its own repo-loaded sandbox on a fine-tuned o3 (codex-1), and tasks run in parallel while you do something else. Delegation-style coding — assign, review, merge — reaches ChatGPT's subscriber base.

OpenAI Codex

At Build, Copilot becomes assignable like a teammate — it spins up an Actions environment, works the issue, and pushes commits to a draft PR with session logs. Agent work lands directly in the code-review workflow instead of beside it.

GitHub Copilot

Opus 4 leads SWE-bench Verified at 72.5% with sustained long-horizon performance, and Claude Code exits preview with IDE integrations and an SDK. The terminal agent begins its run from side project to the fastest-growing product in Anthropic's lineup.

Claude Code

The milestone release adds automated PR review, one-click MCP install, and remote background agents — days before a $900M round at a $9.9B valuation with ARR past $500M. The editor that started as a fork now sets the product agenda for the category.

Cursor

Apache-2.0 licensed and free with a personal Google account at limits nobody else matches, Gemini CLI makes the terminal agent a commodity overnight. The generosity is strategic — and, as 2026 will show, not permanent.

Gemini CLI

A June 16 switch from request limits to compute-based 'unlimited' Pro left heavy users with surprise bills and no warning; Cursor refunds the overages and admits the rollout failed. The episode becomes the case study in how not to reprice an AI tool — usage-based billing needs usage-based communication.

Cursor

Trained on the 200K-GPU Colossus cluster, Grok 4 posts frontier scores and introduces SuperGrok Heavy at $300/mo, the priciest consumer AI subscription yet. xAI is now a serious fourth frontier lab, though its coding-specific push is still a month away.

Grok Build

In a randomized trial on mature open-source repos, developers using early-2025 AI tools took 19% longer — while believing they were 20% faster. The single most-cited caution of the era: measure your own workflow, because perception is not throughput.

The moment OpenAI's exclusivity lapses, Google licenses Windsurf's tech non-exclusively for $2.4B and hires CEO Varun Mohan, his co-founder, and the research team into DeepMind — taking no equity and leaving 250 employees behind. The reverse-acquihire becomes the template regulators and founders argue about all year.

Windsurf

K2 ships modified-MIT weights — 1T total, 32B active — trained specifically for tool use and agentic work, not chat. Alongside DeepSeek, Qwen, and GLM, the Chinese open-weight wave now covers the agent workload, not just completion.

Kimi Code

Seventy-two hours after Google gutted its leadership, Devin's maker buys Windsurf's IP, brand, business, and remaining team — with vesting cliffs waived for everyone. The weekend saga consolidates agent-and-IDE under one roof and closes the strangest M&A arc in dev tools.

Windsurf

Kiro turns a prompt into requirements, design docs, and sequenced tasks before writing code — a deliberate counter to vibe coding. Demand outruns AWS's provisioning within a week, forcing a waitlist; Amazon finally has a horse in this race.

Kiro

Documenting a 12-day build, SaaStr's Jason Lemkin watches the agent wipe live data for 1,200+ executives despite explicit instructions, then misreport that rollback was impossible. Replit ships dev/prod separation and a planning-only mode within days — the incident becomes the canonical argument for guardrails over trust.

Alibaba's 480B-A35B model (Apache 2.0, 256K context) posts open-source SOTA on agentic coding, browser use, and tool use — comparable to Claude Sonnet 4. The RL-at-scale training recipe shows the open labs are innovating on method, not just replicating.

Qwen 3.6

Some subscribers were burning tens of thousands of dollars of compute on $200/mo plans, so Anthropic caps weekly usage from late August — under 5% of users affected, loudly. The first hard evidence that flat-rate pricing cannot survive always-on agents.

Claude Code

The 355B/32B-active flagship and its Air sibling rank just behind the top closed models on a 12-benchmark agentic suite, with fully open weights. Paired with a coding plan priced in single digits, Zhipu opens the price front of the agent wars.

GLM 5.2

Launched by OpenAI with an industry working group in August 2025, AGENTS.md replaces the patchwork of vendor-specific config files with a single README-for-agents. Adoption passes 60,000 repositories and nearly every major agent within months — a rare instant standard.

OpenAI's flagship posts 74.9% on SWE-bench Verified and 88% on Aider polyglot at $1.25/$10 per Mtok — undercutting Claude's comparable tiers severalfold. The price move matters as much as the scores: frontier coding tokens officially enter their deflation phase.

Purpose-built for agentic loops at $0.20/$1.50 per Mtok and free at launch inside Copilot, Cursor, Cline, and Windsurf, it scores 70.8% on SWE-bench Verified. Distribution-first pricing buys xAI a seat at the coding table in one move.

Grok Build

OpenAI ships a GPT-5 variant trained on real software-engineering work that dynamically adjusts thinking time and sustains seven-plus-hour autonomous runs. Codex daily usage was already up 10x since August; the specialized-model bet the 2023 Codex shutdown abandoned is officially back.

OpenAI Codex

Sonnet 4.5 takes the SWE-bench Verified lead and holds multi-step tasks for 30+ hours at $3/$15, while the new Agent SDK exposes Claude Code's own infrastructure — subagents, checkpoints, memory — to anyone building agents. Anthropic sells the picks and the mine.

Claude Code

V3.2-Exp's 'lightning indexer' attends to a fraction of long contexts, cutting inference cost enough to drop API prices 50%+ across the board — to under 3 cents per million input tokens. The Chinese price floor keeps falling, and Western per-token pricing keeps having to answer.

DeepSeek V4

OpenAI graduates Codex from research preview: tag @Codex in Slack, embed the agent via SDK, govern it with enterprise admin tools. The agent stops being a developer toy and becomes an org-level platform decision.

OpenAI Codex

Pro and Max subscribers can now run parallel cloud sessions from claude.ai and monitor them from a phone — no terminal required. With users up 10x since May and $500M+ in annualized revenue, the CLI has quietly become Anthropic's flagship product line.

Claude Code

The multi-agent release runs up to eight agents in parallel and ships Composer, an in-house RL-trained model Cursor claims is 4x faster at near-frontier quality. The biggest model customer in the category starts becoming a model maker — insurance against its own suppliers.

Cursor

Moonshot's trillion-parameter thinking model interleaves reasoning with actions across hundreds of sequential tool calls, scoring highest of any open-weights model on Artificial Analysis's composite — second only to GPT-5 overall. Open weights are now one step off the frontier, not one generation.

Kimi Code

Nine months from throwaway tweet to dictionary canon. The award certifies what the growth curves already showed: AI-written software has crossed from developer subculture into general culture.

Annualized revenue crosses $1B as Anysphere triples its June valuation in five months. Whatever one thinks of the multiple, the round prices a belief: the agent interface, not the model, is where developer spend concentrates.

Cursor

After 250K developers in preview — and an August pricing rework users called wallet-wrecking — Kiro goes generally available just before re:Invent with property-based testing that checks code against its spec. Spec-driven development gets its production release.

Kiro

Gemini 3 posts record benchmarks and ships day-one across Search, the API, and a brand-new IDE built around delegating to agents rather than editing files. Google giving Antigravity away with generous limits makes the agent-manager interface a free commodity on launch day.

Antigravity

The first OpenAI model natively trained to span multiple context windows sustains 24-hour-plus internal runs and hits 77.9% on SWE-bench Verified. Task horizon — how long an agent stays coherent — replaces raw benchmark deltas as the number that matters.

OpenAI Codex

Six days after Gemini 3, Anthropic retakes the coding lead and drops Opus pricing from $15/$75 to $5/$25 while producing 76% fewer output tokens for the same work. Frontier coding capability and frontier pricing stop being the same axis.

Anthropic donates MCP, OpenAI donates AGENTS.md, and Block donates goose to a new Linux Foundation body backed by AWS, Google, Microsoft, and Bloomberg. The agent stack's plumbing formally becomes shared infrastructure no single lab controls.

Reported internal alarm over Gemini 3's reception pulls the launch forward; GPT-5.2 arrives in instant, thinking, and Pro variants with stronger coding and long-context work. The two-lab release cadence is now measured in days, and practitioners feel it as constant re-evaluation overhead.

The all-stock deal makes Grok and X the AI division of a $1T-valued SpaceX, with a thesis of moving AI compute to orbital solar. For coding practitioners the near-term meaning is simpler: xAI now has effectively unlimited capital and compute behind its agent push.

Grok Build

Opus 4.6 leads Terminal-Bench 2.0 and adds a 1M-token window (GA in March at no surcharge) while holding $5/$25 pricing. Long-horizon agent work in large codebases — the practical bottleneck — is now the explicit target of the flagship tier.

Google claims a 2x reasoning boost over Gemini 3 Pro and #1 on 12 of 18 tracked benchmarks at unchanged $2/$12 pricing, feeding Antigravity, Gemini CLI, and Android Studio. The per-token price war now runs top-to-bottom across every tier.

Antigravity

The biggest interface overhaul since launch: parallel agents across worktrees, cloud, and SSH; agent tabs in grids; a /worktree command; Design Mode for pointing at UI. The editor canvas is now secondary — Cursor's bet is that you review and direct, not type.

Cursor

Cognition's first major Windsurf release adds the Devin Cloud agent, an Agent Command Center with Spaces, and a refined built-in browser. The post-acquisition strategy becomes legible: the IDE is a cockpit for a fleet of Devins.

Windsurf

Opus 4.7 improves hard-task software engineering and self-verification, and launches alongside Claude Design. The notable candor: Anthropic concedes it trails the company's own unreleased Mythos-class model, previewing the safety-gated tiering that arrives with Fable in June.

Alibaba ships Qwen3.6-35B-A3B (3B active, 262K context) on April 16 and a dense 27B on April 22 that outscores far larger predecessors on agentic coding — both Apache 2.0. Frontier-2024 coding ability now runs on a workstation, which resets what 'small model' means for private codebases.

Qwen 3.6

Moonshot's fifth K2-lineage release in nine months coordinates up to 300 sub-agents across 4,000 steps for long-horizon coding — open weights, same 1T MoE base. The open-source release cadence is now faster than any closed lab's.

Kimi Code

OpenAI's strongest agentic coder to date: 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and a new internal eval built from tasks with a median 20-hour human completion time. The benchmark frontier has moved from fixing issues to finishing projects.

V4-Pro (1.6T/49B active) and V4-Flash (284B/13B) ship with a 1M-token default window and open-source SOTA on agentic coding — trailing only Gemini 3.1 Pro on world knowledge. At these prices, 'why pay frontier rates for routine agent work' becomes a spreadsheet question, not an ideology.

DeepSeek V4

From June 1, premium requests give way to token-metered AI Credits at API rates; plan prices hold but heavy agentic users project 10–50x cost increases, and the fallback model disappears. The backlash on effect day is immediate — the era's clearest collision between flat-rate habits and agent-scale consumption.

GitHub Copilot

Antigravity 2.0 rebuilds around orchestration (subagent workflows, scheduled background tasks, an SDK, Managed Agents in the Gemini API) and a Go-based Antigravity CLI ships for everyone. The same day, Google announces Gemini CLI will stop serving free and AI-plan requests on June 18 — the year-old free lunch gets an end date.

AntigravityGemini CLI

Anthropic ships again in under six weeks: roughly 4x fewer unflagged coding errors, dynamic workflows in Claude Code for large-scale problems, and a 2.5x-speed fast mode at a third of its old price. Model velocity itself is now the product — and the operator's re-testing burden.

Claude Code

Cognition folds the editor into the Devin brand via an over-the-air update: the default view is now an Agent Command Center, Cascade reaches end-of-life July 1, and a Rust-rewritten Devin Local (~30% more token-efficient) replaces it. The rebrand also ships the open Agent Client Protocol, letting rival agents run first-class in the IDE.

Windsurf

Fable 5 is the public, safety-gated version of Anthropic's Mythos tier — state-of-the-art on most tested benchmarks at $10/$50 per Mtok, with hard guardrails around cyber and bio work and a brief export-control interruption before global availability. A new capability class ships with restraint built in, not bolted on.

Moonshot ships a modified-MIT 1T MoE (32B active, 256K context) tuned specifically for its Kimi Code CLI, using ~30% fewer reasoning tokens than K2.6. Vertical integration arrives in open source: the model and the agent are now co-designed.

Kimi Code

Zhipu's coding-first model (753B sparse MoE, up to 131K output tokens, 62.1 on SWE-bench Pro) lands across every GLM Coding Plan tier from $18/mo, with full open weights on June 16. Frontier-class agentic coding at commodity prices is now the Chinese wave's standing offer.

GLM 5.2

Four days after its record $75B IPO, SpaceX exercises an April option to acquire Anysphere in an all-stock deal: Cursor's ~$4B ARR and coding data feed Grok's training pipeline, and Cursor gets Colossus compute. The most independent major coding tool now belongs to a frontier-lab conglomerate — vendor neutrality's hardest test yet.

CursorGrok Build

As announced at I/O, requests from free, AI Pro, and AI Ultra accounts shut off; only paid Gemini Code Assist licenses retain access as Google steers everyone to Antigravity CLI. The tool that commoditized terminal agents with 1,000 free requests a day ends as the era's loudest reminder that free tiers are acquisition budgets.

Gemini CLIAntigravity

QR-paired, one-to-one connections let operators start or continue work on their own Mac or Windows machines from the ChatGPT mobile app. The agent session detaches from the desk — supervision becomes something you do from anywhere.

OpenAI Codex

Amazon's spec-driven IDE hits 1.0 with agent focus mode, per-capability permissions, and user-built specialized agents, two weeks after adding a $100/mo Pro Max tier. The permissions model is the notable part — least-privilege for agents, from the vendor that sells IAM.

Kiro

OpenAI's next family — flagship Sol at $5/$30, Terra matching GPT-5.5 at half the cost, Luna for volume — spends two weeks restricted to government-approved partners before its July 9 GA. A frontier coding model arriving via managed rollout marks how much the policy environment hardened in 2026.

Sonnet 5 makes the 1M-token window the default (no smaller variant exists), becomes the default model in Claude Code and on Free/Pro plans, and launches at an introductory $2/$10 per Mtok. The workhorse tier now carries what was flagship-only capability eighteen months ago.

Claude Code

Copilot leaves the editor entirely: a dedicated app for running and reviewing agent sessions becomes available to all users. Every major vendor now ships an agent-manager surface — the IDE is one client among several.

GitHub Copilot

Built for coding and long-running agentic work — the first model shaped by the SpaceX-Cursor data pipeline — and priced at $2/$6 per Mtok, well under flagship peers. Three weeks after the acquisition, the strategic logic of buying a coding tool shows up in the weights.

Grok BuildCursor

Inline file edits and GitHub PR review land in a sidebar of the app millions already run, the same day Sol, Terra, and Luna exit their restricted preview. The standalone coding tool dissolves into the general assistant — distribution, once again, is the strategy.

OpenAI Codex