Source-linked historical snapshots
Today's comparison uses November 8, 2026. Each year resolves to the latest documented frontier state available on that same day—not a fresh daily leaderboard. Tools and specialist models appear only after their release. Rankings are editorial; the limits are the point.
Last year
The frontier stopped being just a chat window. Reasoning, tools, terminals, and multimodal generation were becoming one workflow.
Lineup last changed · Claude 4 launch
What each could do—and what it still could not.
Anthropic
Known for
Anthropic positioned it for hard, tool-using repo tasks—not for unattended product delivery.
Could
Given a well-scoped issue, a repo, and terminal tools, inspect the code, change multiple files, run tests, and return a serious candidate patch.
Couldn't
Choose the right product behavior from vague intent, notice constraints outside the repo, or earn a merge without a human reviewing the diff and the test coverage.
Anthropic
Known for
A stronger, more instruction-following sibling to Opus 4 for agent and coding workflows.
Could
Follow detailed coding instructions, use tools in parallel, and complete more of a defined engineering loop than prior Sonnet models.
Couldn't
Turn a weak specification into the right product, know hidden operational constraints, or make a production change safe without human review.
Google DeepMind
Known for
Strong at turning a detailed prompt into an impressive interactive app or large-context analysis.
Could
Turn a detailed prompt into a polished interactive web demo, transform code, or pull relevant facts from a very large supplied context.
Couldn't
Prove the demo handled real edge cases, discover an unstated product requirement, or keep an open-ended engineering project on track by itself.
OpenAI
Known for
The model to combine web search, files, Python, images, and deliberate reasoning in one answer.
Could
Search, read files, write and run Python, inspect images, and synthesize the results into a careful answer or bounded technical plan.
Couldn't
Reliably absorb a team’s unwritten conventions, make product tradeoffs, or operate a repository through many handoffs without someone steering and checking it.
Where models could work.
Terminal-native delegation
You could hand a real repository to an agent, let it inspect files, make edits, and move a substantial task forward.
Anthropic, Feb 24Frontier reasoning in the terminal
An open-source experiment connecting local code and screenshots to reasoning models.
OpenAI, Apr 16A generous open command line
Gemini 2.5 Pro came to the terminal in an open-source agent with broad free access.
Google, Jun 25The specialist state of the art.
Voice model
When generated speech got direction
Emotion tags, multi-speaker dialogue, 70+ languages, sighs, whispers, and more believable turns. Cinematic output, not yet the low-latency default for live conversation.
ElevenLabs, Jun 3Image model
The prompt became a conversation
Text in images, edits, visual reference, and iterative art direction felt native to chat.
OpenAI, Mar 252 years ago
Computer use made agents feel tangible, but it was explicitly experimental: slow, brittle, and nowhere near an unattended employee.
Lineup last changed · Claude computer use
What each could do—and what it still could not.
Anthropic
Known for
An upgraded Sonnet paired with experimental screen, mouse, and keyboard control.
Could
Use experimental computer control to look at a screen, click, type, and carry out a narrow browser or desktop task.
Couldn't
Operate a computer quickly or reliably enough for unattended work; Anthropic reported only 14.9% on OSWorld against a 70–75% human baseline.
OpenAI
Known for
It spent more time on difficult math, science, and coding prompts before answering.
Could
Spend extra compute on difficult math, coding, and science prompts, often catching steps a fast chat model missed.
Couldn't
Use the broad ChatGPT tool set at launch, answer quickly, or make a vague real-world project well-specified on its own.
Google DeepMind
Known for
It could hold a long document, codebase, or video corpus in one unusually large prompt.
Could
Accept an unusually long codebase, book, or video transcript and retrieve, summarize, or compare material from inside it.
Couldn't
Turn that giant prompt into a reliable sequence of edits and validations; seeing the whole context did not make it an autonomous worker.
Meta
Known for
A 405B open-weight model that gave teams a credible self-hosted option for demanding text tasks.
Could
Give technically capable teams a large open-weight model for self-hosted chat, retrieval, and code assistance.
Couldn't
Remove the operational cost of serving and evaluating a huge model, or turn open weights into a reliable agent without a carefully built harness.
Where models could work.
The AI-native editor takeoff
Ask an assistant in the editor, accept a diff, and keep your own hands on the wheel.
Cursor’s storyThe autonomous engineer demo
A browser, terminal, tests, and a progress loop around the idea of a task-owning software agent.
Cognition, Mar 12Plan before the pull request
A path from issue to plan to code—before terminal agents became the default mental model.
GitHub, Apr 29The specialist state of the art.
Voice model
Natural speech was already a product
Narration, localization, accessibility, and creator work were viable; a real-time agent stack was still taking shape. It was polished generation, not a natural live conversation system.
ElevenLabs, Aug. 22 2023Image model
The aesthetic standard
Striking images, better prompt accuracy, coherence, remixing, and long-input handling made it the creator reference point.
Midjourney version history3 years ago
Open weights became a serious alternative. The best systems were still prompted assistants, but teams could now adapt a capable model inside their own stack.
Lineup last changed · Llama 2 launch
What each could do—and what it still could not.
OpenAI
Known for
The reference point for hard prompted reasoning, sophisticated writing, and one-off code help.
Could
Explain a hard concept, repair a small code snippet, draft a strong document, or reason through a bounded problem after a careful prompt.
Couldn't
Read your repo, execute code, inspect files, retain a task state, or verify its answer; you copied the output and did the rest.
Anthropic
Known for
A place to paste a long text, ask questions about it, and get a more useful than usual written response.
Could
Read a long pasted document, summarize it, answer questions about it, and help draft or revise a response in the same text session.
Couldn't
See a screenshot, browse current information, manipulate files, or take a coding task from an issue to a tested change.
Known for
General chat, multilingual writing, and code snippets across Google’s emerging AI products.
Could
Translate, draft, answer general questions, and generate small code snippets across Google’s chat and developer surfaces.
Couldn't
Reliably solve hard multi-step reasoning problems, keep track of a codebase, or compete with GPT-4 for the hardest prompted work.
Meta
Known for
A commercially usable open-weight model that made self-hosted assistants and fine-tuning more practical.
Could
Run in a company-controlled environment for drafting, retrieval, and fine-tuned text assistants without sending prompts to a closed API.
Couldn't
Match GPT-4 at the hardest reasoning tasks, see images, or make an autonomous workflow reliable without substantial engineering around it.
Where models could work.
The first widely felt tool loop
ChatGPT could run Python against uploaded files to analyze data, make charts, edit files, and calculate instead of only describing what to do.
OpenAI, Jul 6The autonomous-agent provocation
It popularized the loop of task, plan, tool, result, and next step—but was fragile, expensive, and far from a reliable worker.
Project repositoryThe first common plumbing
Developers began wiring models to search, calculators, code, and databases, usually with substantial prompt engineering and supervision.
LangChain documentationThe specialist state of the art.
Voice model
Voice cloning went multilingual
High-quality synthetic speech and voice cloning in eight languages opened narration, localization, and creator workflows. It was generated audio, not a natural real-time conversation partner.
ElevenLabs, Apr 27Image model
The aesthetic frontier
It produced striking, stylized images from text prompts and became a creative default. Precise edits, dependable text, and repeatable art direction were still difficult.
Midjourney version history