Best AI Coding Assistants 2026

Claude Code holds first at 9.5, and this week's advice is for the teams running OpenAI Codex at number two: you have a migration deadline.

OpenAI's changelog, dated 14 September, says GPT-5.5 retires from Codex on 14 October for everyone signed in with ChatGPT. The instruction is specific. Switch from gpt-5.5 to gpt-5.6-sol, then update workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and any script that still selects gpt-5.5. The OpenAI API keeps the older model, so the risk sits with people who drive Codex through a ChatGPT login.

Here is how I would run it. Search your repositories and CI configuration for the string gpt-5.5 today. Change one pipeline first, compare a week of its pull requests with the old baseline, then roll the rest. Codex keeps 9.3 overall and 9.5 on agentic capability, and a planned switch protects that quality.

The command line tools shipped steadily this week too. GitHub Copilot CLI 1.0.85 arrived on 16 September with Vim mode for everyone plus new /settings and /config commands. Antigravity CLI 1.2.4 on the same day reloads skills without restarting a session, and Gemini CLI 0.60.0 on 15 September tightened how extensions reach your system.

That cadence is why I pin tool versions in shared environments. A weekly release train brings real improvements, and it rewards teams that upgrade on a schedule they choose.

Claude Code stays first on the strength of code quality at 9.8 and agentic capability at 9.7, and its release on 10 September added unified pricing across Claude apps. Cursor at 9.1 keeps IDE integration at 9.6, the best editor experience here. GitHub Copilot at 7.9 holds 9.5 on IDE integration, and for a team that lives in VS Code and wants the lightest setup, that number decides it.

The ranking carries forward unchanged.

Claude Code holds first at 9.5, and this week the story is the bill at the end of a long session.

Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on 1 September at the same headline rates, $10 per million input tokens and $50 per million output. The cache is what changed. Cache reads dropped from $1 to $0.25 per million, a 75 percent cut, and cache hits now bill at 0.025 times the base rate. Anthropic puts the effect at roughly 25 percent off a typical workload and up to 45 percent off an agentic one.

Agentic coding is the workload that gains most, because an agent rereads the same repository map, the same file headers and the same instructions on every turn. A four hour refactor that reads a large cached context hundreds of times now costs a fraction of what it did in August. I raised Claude Code's value score to 9.2 for that reason, and it keeps first place overall at 9.5.

The practical move is structuring your prompts so the cache actually hits. Put the stable material first, your project instructions, your architecture notes, your style guide, and let the changing part of the conversation come last. A cache that breaks on every turn gives you none of this discount.

OpenAI Codex holds 9.3 and stays the strongest option for teams already living inside the OpenAI stack, with agentic capability at 9.5. Cursor at 9.1 keeps the best editor experience in the category, scoring 9.6 on IDE integration, and it is what I hand a developer who wants an agent without leaving the editor.

Aider at 7.7 stays the cheapest serious path, with a value score of 9.5. It is open source, terminal native, and free to point at whichever model is cheapest this month, which makes it a useful companion to any of the hosted tools above.

Claude Code keeps first at 9.5. This week's most useful lesson for coding teams came from the September 3 outage, which reached the tools developers lean on hardest. OpenAI said a routing error made ChatGPT and Codex unavailable for some users for roughly half an hour, and Anthropic reported a partial infrastructure outage covering Claude Code and its developer API. A sprint that depends on one hosted agent stalls the moment that agent does.

My response is a written fallback plan. Keep a second agent configured on your main repository and practice switching to it before you need it. Aider at 7.7 is my favourite spare: it is open source, runs in any terminal, and connects to whichever model provider is up, including a local model on your own machine.

Claude Code stays on top for long multi-file refactors where the test suite defines done. It plans the change, edits in careful steps, and runs the tests between them, and that discipline is why I trust it with the biggest jobs.

OpenAI Codex holds 9.3 while I keep pushing the Agents API public beta through real repositories. Cursor at 9.1 remains the best editor for reviewing agent diffs. Cursor is now owned by SpaceX and has added Origin code hosting and longer-running cloud agents, two features I am testing before its score moves.

Pricing has settled into clear tiers. As of September 1, GitHub Copilot Pro is the cheapest paid entry at $10 a month, Claude Code, Codex, and Cursor start around $20, and daily agent work realistically costs $60 to $200. Copilot's CLI, generally available since February, bundles a cloud agent and a code review agent, which makes the $10 plan a strong starting point for students and hobbyists.

Google Antigravity at 8.5 and Jules at 8.2 remain easy picks for teams already living in Google's tools. No rank changes this week.

Claude Code keeps first at 9.5, and this week's news made the top three more interesting without reordering it. The headline event was OpenAI's September 10 launch of the Agents API in public beta. It brings the Codex harness to cloud agents with managed orchestration, long-running sessions, and OpenAI-hosted sandboxes, and Deep Research also landed inside Codex. For teams that want coding agents running in their own pipelines, that is a real expansion of what OpenAI Codex can do.

I held Codex at 9.3 for now. A public beta earns score once teams have pushed it through real repositories for a few weeks, and I will be testing it hard before the next update.

Claude Code stays on top for the work that fills most of my week: long refactors across many files with the test suite as the contract. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 on September 1, and Claude Code keeps landing steady maintenance releases. Version 2.1.270 fixed a regression where read-only git commands asked for permission in long sessions. Small fixes like that matter, because permission prompts are exactly where an agent's momentum stalls.

Cursor holds 9.1 as the best editor experience for developers who want an agent inside the IDE, with fast tab completion and a clean diff review surface. The New Stack described Cursor, Claude Code, and Codex settling into a composable stack of orchestration, execution, and review layers. That matches how I work: the heavy task runs in a terminal agent, and the review happens in the editor.

Check pricing before you commit. GitHub Copilot Pro remains the cheapest paid entry at $10 a month, most agent tools start around $20, and daily heavy agent use pushes a realistic budget toward $60 to $200 a month.

Google Antigravity stays fourth at 8.5 and Jules fifth at 8.2. Jules comes bundled with Google AI Pro and Antigravity is free to start, which makes both easy recommendations for anyone already living in Google's tools. No rank changes this week.

Ten days after the September model wave, I have run enough real repository work through these tools to say the ranking holds and to explain what is actually deciding it. Claude Code stays first at 9.5 on agentic reliability across long tasks. Hand it a refactor spanning a dozen files with the test suite as the contract, and it works the whole job, reads what it needs, runs the tests, and tells you plainly when something breaks. That last behaviour is worth more than any benchmark line, because an agent that reports success on a red suite hands back the hour you thought you had saved.

OpenAI Codex holds 9.3 on the cloud sandbox story. Sending a long job to run remotely while you keep typing locally pays off the moment your repository grows large enough that a full test run breaks your concentration.

Cursor keeps 9.1 as the sharpest pure editing experience. Completion that understands the surrounding project and inline edits that land where you expect still make it the fastest tool for an hour when you already know what you want to write.

The development I keep watching this month is price. Anthropic cut cache-read pricing by three quarters at the start of September, and that matters here more than in any other category, because coding agents re-read the same files dozens of times inside one session. A heavy working day now costs measurably less than it did in August. Choose one tool, learn how it decides what to read, and go deep. Depth in a single assistant beats a rotating trial of three.

GPT-6 Astra landing on September 3 is the first thing in months that has made me re-examine the top of this list, and after a week with it I am keeping Claude Code at number one while nudging OpenAI Codex up to 9.3. Astra is a real step up in multi-file reasoning, and Codex inherits that immediately. What it does not inherit is the harness quality that keeps Claude Code ahead: the permission model, the way it holds a long refactor together across a repo, and the fact that it recovers gracefully when a test run fails midway.

That distinction matters more than model benchmarks now. Every serious assistant in this list is wired to a frontier model, so the differentiation has moved to how the agent behaves over a forty-minute task. Claude Code still handles that stretch best in my daily use.

Cursor holds third at 9.1. The editor experience remains the smoothest way to work if you want to stay inside an IDE, and for developers who think in terms of open buffers rather than terminal sessions it is the right pick.

One practical note on cost. Astra's output pricing is steep enough that agentic loops which used to run cheaply on the previous generation now add up fast. If your assistant lets you set a model per task, put the frontier model on architecture and debugging, and let a fast tier handle mechanical edits and test scaffolding. I have cut my own weekly spend by roughly a third doing exactly that.

Two things happened this week that matter for anyone paying for a coding assistant. Anthropic shipped a 1M token context tier for its agent line and cut prompt cache read pricing by seventy five percent. Both land directly on the economics of agentic coding, which is the workload where these tools either earn their subscription or quietly waste it.

The cache pricing cut is the bigger deal and it gets less attention than it deserves. Agentic coding sessions re-read the same repository context on every turn. When cache reads cost seventy five percent less, a long refactoring session that used to burn through a monthly allowance in two afternoons now stretches across a week. I raise Claude Code's context window score to 9.7 for the combination of the larger window and the pricing that makes actually using it viable.

OpenAI Codex moves to 9.2. The GPT-6 Astra rollout is gated, and the reason I am adjusting Codex upward is separate: its multi-file editing has become noticeably more reliable at holding an interface contract across the files it touches. I gave it a task this week that required changing a function signature in one module and updating eleven call sites, and it got all eleven.

Cursor holds at 9.1 and remains my recommendation for developers who want the assistant inside a familiar editor. A McKinsey survey published this cycle found that thirty two percent of organizations have skipped buying a software product because agentic coding tools let them build it internally. That number tells me the category has moved past experimentation, and it is why I weight agentic capability so heavily in this ranking.

Two frontier releases in five days and the ranking barely moves, which tells you something important about this category. Claude Code holds first at 9.5 because the product is the harness, and the harness is what determines whether a coding agent finishes a task or leaves you a half-refactored repo. Anthropic's Fable 5.1 on September 1 was tuned specifically for agentic tool use and shipped with cache reads at $0.25, and long autonomous sessions are exactly where that pricing lands. OpenAI Codex at 9.1 gets the more interesting week, because GPT-6 Astra reports 57.7 on Terminal-Bench 4.0 and 74.1 on DeepSWE, which are the two benchmarks that actually correlate with finishing real engineering tickets. I want to see those numbers hold up in a working repo before I move Codex above 9.1, and I expect to revisit this within a fortnight. Cursor stays at 9.1 as the best answer for developers who want an editor first and an agent second, and the multi-file edit flow remains the most legible in the category. Google Antigravity at 8.5 is improving faster than anything else on this list. My honest advice: the model underneath these tools now changes every few weeks, so pick your assistant on how well its harness recovers from a failed step, because that is the property that does not churn.

Anthropic released Claude Fable 5.1 on September 1, calling it their most advanced model for coding and knowledge work, explicitly tuned for agentic tool use and shipped with much cheaper cache reads. That last detail is the one that matters for this category, and it is why Claude Code holds first at 9.5. Long agentic coding sessions replay enormous amounts of context on every turn, and the cost of that replay is what has kept teams from letting an agent work unsupervised for an hour. Cutting cache read pricing changes the economics of the whole workflow. Claude Code stays first on the same three strengths I have argued all year: it reads a large repository properly before editing, its edits survive code review, and it recovers from its own mistakes without a human restarting the session. OpenAI Codex at 9.1 remains the better pick inside the ChatGPT ecosystem and its cloud task execution is genuinely useful for parallel work. Cursor at 9.1 is still the best pure editor experience, and for a developer who wants tab completion and inline diffs over a terminal agent, it is the right tool. Google Antigravity at 8.5 has improved fast and its browser-aware debugging is a real differentiator for front-end work. Jules at 8.2 is the free option worth using. My verdict for September: pick your assistant by where you want the agent to live. Terminal-first developers should be on Claude Code, editor-first developers on Cursor, and everyone should be checking their monthly token spend now that long sessions are affordable.

Claude Code stays at the top of this list and August did nothing to challenge it. The month's sixteen model releases across eight providers were overwhelmingly speed and cost tiers, which is a useful thing for coding agents but does not move the ceiling on what an agent can finish unattended. That ceiling is where Claude Code earns its 9.5. Opus 5 has been the underlying model since late July at roughly half the previous flagship cost, and the practical effect is that long agentic sessions, the kind where the tool reads twenty files, runs a test suite, and iterates, became affordable to run repeatedly, all day long. OpenAI Codex and Cursor stay tied at 9.1 for different reasons. Codex has the cleanest handoff between a chat conversation and a running task, and Cursor still owns the in-editor experience for developers who want to stay in the file and steer. Google Antigravity holds fourth. Gemini 3.7 Flash landing on August 13 as the fastest frontier-class model helps Antigravity specifically, because a lot of what an IDE agent does is many small cheap calls where latency compounds. I keep the ranks unchanged this week because none of these releases changed what I would actually reach for. The advice stands: pick Claude Code if you want an agent that finishes multi-file work, Cursor if you want to stay in the editor, and Codex if your team already lives in the OpenAI stack.