Best AI Coding Assistants in 2026: Pick the Job, Not the Brand
The best AI coding assistant in 2026 depends on the job: Cursor or Trae for IDE-native writing, Claude Code (Opus 5.5 and Sonnet 5.5 by default) for terminal-driven agentic work, Codex for delegating tasks to the cloud, and Devin for fully autonomous tickets. Cursor's plans now run $0, $8, $30, $100 and $200 a month.
The short version
AI coding assistants still split into five jobs, not one category, but the details underneath moved again: Cursor now sells through The X Subscription from $8 a month, Claude Code defaults to the 5.5 models, Codex added GPT-6.1 Terra and priced Ultrafast, and Tabnine is contact-sales only. Pick the job first, then re-check the price.
Not sure which one fits you? Describe your situation and Smart Match names three tools.Find my AI tools
In this guide19 sections
Eighty-four percent of developers now use or plan to use an AI coding tool. Only 29% of them trust what it produces, down from roughly 40% two years earlier, per Stack Overflow's 2025 Developer Survey. That gap didn't open because the models got worse. It opened because "AI coding assistant" stopped meaning one thing while most roundups kept ranking it like it still did.
For a four-to-twenty person engineering team, that matters more than any leaderboard. These tools now split across five genuinely different jobs: writing new code, running whole tickets unattended, searching a codebase, reviewing pull requests, and fixing what's already live in production.
The biggest structural move in the category this year came from a company doing the opposite of specializing. On June 2, 2026, Cognition folded Windsurf into its Devin product line, consolidating two jobs under one brand rather than sharpening one. Picking a tool without picking a job first is how a team ends up paying for three assistants that do the same thing and nothing that covers the other four.
The tools on this shortlist are mostly the same five weeks on. What is underneath them is not. Claude Opus 5.5 became the default model inside both Claude Code and GitHub Copilot within the same 24 hours in late September, and one name from the original lineup, Tabnine, changed hands entirely.
Cursor's cloud agents can now run on a team's own machines instead of Cursor's, and even the models outside this shortlist moved: OpenAI cut prices roughly in half on its own coding-capable pair, GPT-6 Sol and GPT-6 Luna, the same week. A framework built around jobs survives a model refresh better than one built around a leaderboard, but the specifics below are the ones that moved.
Reviewing pull requests is a separate purchase with its own pricing, covered in the best AI code review tools.
Hokai chart of the guide's shortlist by job, first paid plans from each vendor's pricing page as checked between August and November 2026.
What changed since 2 October 2026
Re-checked on 6 November 2026 against each vendor's own pricing page. The five jobs and the picks below hold; the prices and model lineups underneath them moved:
- Cursor changed how it sells. New customers now buy a bundled subscription that replaced the old Hobby, Pro and Pro+ ladder; existing subscribers keep their legacy plan.
- Claude Code's default models moved to the 5.5 generation, and its top plan was renamed.
- Codex gained a cheaper model and a priced speed mode. A cheaper model joined its usage table, and OpenAI now publishes what Ultrafast mode costs.
- GitHub Copilot now meters agent work in monthly AI credits. The four plan prices did not move.
- Greptile dropped per-seat pricing, and a new Google entrant joins the "not on the shortlist" section.
- Unchanged on re-check: Devin Desktop, Trae and Tabnine's contact-only sale.
What "AI coding assistant" actually means in 2026
Five jobs, five different buying decisions:
- Writing new code inside an editor. Cursor and Trae live here: fast, IDE-native, built around autocomplete and inline agent requests.
- Running a whole task from a terminal, mostly unattended. Claude Code is the default pick; it reads a repo, edits files, runs commands and iterates without a human approving every step.
- Handing off an entire ticket end to end. Devin and its IDE sibling Windsurf (rebranded Devin Desktop on June 2, 2026) sit at the fully-autonomous end, closer to a contractor than a plugin.
- Searching and understanding a large, unfamiliar codebase. Sourcegraph does this, and only this, at enterprise scale.
- Reviewing pull requests and debugging what's already shipped. Greptile reviews PRs automatically; Lightrun lets a team inspect a live production process without redeploying it. Neither writes a line of new code.
Most "best AI coding assistant" roundups still rank all of these against each other on one list: a widely-read coding-tools roundup puts an editor plugin, an app builder and an enterprise search product in the same "best use case" table as if they were interchangeable.
Treating the category as one leaderboard is why so many teams end up owning three tools that do the same job and nothing that covers the other four. Hokai's own AI coding assistants directory lists more than a dozen entries; almost none of them compete directly with each other once you sort by job instead of by name.
How to choose: four questions before you compare tools
1. What surface does the work actually happen on? If most of the team lives in an IDE, an editor-native tool wins on friction alone. If the heaviest work is repo-wide refactors kicked off from CI or a terminal, an agentic CLI tool fits better regardless of how good the IDE competitor's autocomplete is.
2. Does anyone need to hand off a ticket completely, or does an engineer stay in the loop? Devin and Devin Desktop are built for the former: genuinely autonomous, billed by compute consumed rather than seats. Cursor, Claude Code, Trae and Tabnine assume an engineer is reviewing every diff.
3. Is the codebase the bottleneck, or the review queue, or production itself? A team that can already write code fast but drowns in PR review needs a review tool, not a second code-generation tool. A team debugging incidents in a live service needs runtime visibility, which none of the writing-focused tools provide.
4. Who is actually approving the purchase? Under about ten engineers, this is usually the engineering lead and a company card. Above that, it usually becomes a security review, and self-hosting or compliance certifications start mattering more than any per-seat price. That fourth question is the one teams skip, and it's the one that most often overturns the other three.
Still not sure which job is the actual bottleneck? Smart Match asks a version of these same four questions and narrows Hokai's directory to a short list instead of asking a team to work through this alone.
What each tool actually costs
- Cursor: a Free tier, then four paid plans from $8 to $200 a month (detailed below); annual billing takes the price to $80, $300, $1,000 and $2,000 a year. Subscribers who joined before the change keep the legacy Hobby, Pro at $20, Pro+ at $60 and Ultra plans.
- Claude Code: no free tier at all; Pro $20/month ($17/month billed annually); Max $100/month; Ultra $200/month for the heaviest use. Team standard seats are $20 a month billed annually ($25 monthly) and premium seats $100 ($125 monthly).
- OpenAI Codex: no separate subscription. It rides on a ChatGPT plan: Free ($0) and Go ($8/month) on the lighter GPT-6 Luna model in the desktop app, Plus $20/month, Pro at $100, $200 or $500/month, Business $20 per user per month billed annually ($25 billed monthly), and custom Enterprise and Edu pricing. Usage past a plan's included limits draws on purchased credits. OpenAI lists Ultrafast mode access on the $500 Pro plan.
- Windsurf (Devin Desktop): Free; Pro $20/month; Max $200/month; Teams from $80/month minimum, either $40/month per full seat or unlimited shared-credit "flex" seats.
- Devin: the same ladder as Devin Desktop above, since the two now share one billing page.
- Tabnine: contact sales only. Tricentis's July 2026 acquisition took Tabnine's self-serve pricing page offline; there is no published per-seat price left to quote.
- Trae: free for 5,000 completions a month with no card required, then a paid ladder rebuilt this cycle to charge exactly what Cursor charges: $20, $60 and $200 a month. The old $3/month Lite tier no longer exists.
- GitHub Copilot: Free with 2,000 completions a month; Pro $10/month with $15 in monthly AI credits; Pro+ $39/month with $70; Max $100/month with $200. The cheapest paid seat in this whole list, though agent mode, code review and the CLI now draw on those credits.
- Google Antigravity: $0 for individuals with basic weekly rate limits and unlimited Tab completions; Google's paid AI plans raise the limits and add a credit pool; organisations pay through Google Cloud.
The shortlist, by job
| Bottleneck | Pick | Starting price | The one caveat |
|---|---|---|---|
| IDE-native writing, model choice | Cursor | Free; Plus $8/mo; Premium $30/mo | Third-party models need Premium or above and bill at API list price |
| Terminal-first, deep repo work | Claude Code | $20/mo (Pro), bundled with Claude | No free tier at all; defaults now Opus 5.5 and Sonnet 5.5 |
| Delegate a task to the cloud, review the result | OpenAI Codex | Free; Go $8/mo; Plus $20/mo | Bundled in chat plans; Ultrafast on $500 Pro |
| Fully autonomous IDE, budget-unified | Windsurf (Devin Desktop) | Free; Pro $20/mo; Max $200/mo | The Windsurf brand and its old pricing page are gone |
| Hand off a whole ticket | Devin | Free; Pro $20/mo; Teams from $80/mo | No longer priced per Agent Compute Unit for self-serve users |
| Regulated or self-hosted teams | Tabnine | Contact sales | Tricentis acquired Tabnine in July 2026; self-serve pricing is gone |
| Fits inside a free tier | Trae | Free (5,000 completions/mo) | Paid tiers copy Cursor's old ladder |
Cursor and its new plan ladder
Cursor. The biggest change since the last review is how Cursor is sold. Its help centre now says new customers buy a bundled subscription that covers Grok, Cursor and X, and choose a plan there. Every paid plan includes a weekly usage pool for Cursor's own models, shared with Grok.
Read from the vendor's pricing page; Free is $0. Super is 15x Plus usage and Ultra 50x.
The catch is what each rung unlocks. The cheapest paid plan runs Cursor's own models at fixed settings and leaves out Auto, cloud agents, automations, automated code review and third-party models. Premium and above add all of that, with outside models charged at API list price from prepaid credit rather than from the plan.
Since September 2, 2026, Cursor's cloud agents can also run on self-hosted machines: a single connected laptop for personal use, or a pooled queue of a team's own workers, so a codebase, its build outputs and its secrets never have to leave a company's own network. That closes some of the distance to Tabnine's compliance pitch below, though Cursor still isn't sold as a zero-retention product the way Tabnine was.
Claude Code and the 5.5 models
Claude Code. Ships inside Anthropic's Claude subscription rather than as a standalone product, which is the single biggest reason teams pair it with a free-tier IDE tool rather than starting there. What changed: Claude Opus 5.5, released September 22, 2026, is now the default model across Claude Code, the Claude app and Cowork for Pro, Max and Team users, cutting execution cost roughly 40% against the previous Opus 5.
The model menu has since widened. According to Claude Code's own model documentation, on the Anthropic API the opus, sonnet and haiku aliases now resolve to Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5.
Anthropic's pricing page lists Sonnet 5.5 at half the API price of Opus 5.5, so Sonnet is the model to try first for well-scoped tickets. The most capable model is opt-in through /model fable, and on some plans it bills to usage credits instead of included limits. The top plan is now called Ultra and, for verified accounts, has no five-hour limit.
Weekly usage limits also moved in mid-September. Anthropic's "permanent 25% increase," effective September 14, replaced a temporary 50% boost that had been running since May, so most teams' real weekly ceiling is lower than it was the week before, not higher, despite the headline framing. The separate weekly cap that used to apply specifically to Opus usage was removed the same day.
OpenAI Codex
OpenAI Codex. Codex is the delegate-and-review option on this list: describe a task, let it work in an isolated environment, then read the diff. It has no price of its own and comes with the ChatGPT plan a person already pays for, so the real cost question is which plan's usage limits a team will hit, not which tool to buy.
Its engine is GPT-6 Sol for multi-step engineering work and GPT-6 Luna for small, high-volume edits. Since September 29, 2026 it also runs GPT-6.1 Sol, which OpenAI says delivers near-flagship intelligence at about a fifth of the flagship's standard token prices; that is a vendor claim, not an independent test.
OpenAI's DevDay on September 29, 2026 changed what Codex is for. Cloud environments now let a team start tasks from a shared, pre-approved setup, and tasks can run remotely from a phone or any device. The refreshed command-line tool adds voice control and an /agents view for tracking several delegated tasks at once.
A new Code Review experience in the ChatGPT desktop app summarizes a diff and lets a person question Codex before commenting on GitHub pull requests or GitLab merge requests. Automatic reviews let Codex take a first pass in the cloud while the developer is away, and Codex Security, the vulnerability-scanning agent, gained a cloud version that scans whole GitHub repositories on demand or on a schedule.
Where it loses: the computer-control feature works on macOS only, there is no self-hosted option, and every change still needs a human reviewer.
A team that wants to stay hands-on inside an editor will be better served by Cursor above.
Codex speed modes and models
Ultrafast is the speed mode to price carefully. OpenAI's pricing page says it consumes 8x the included subscription usage of Standard mode on its two top Codex models, or 6x when paid for with purchased credits, and lists access on the $500 Pro plan. The ordinary Fast mode costs 2.5x included usage, or 2x on credits.
OpenAI's usage table also now lists GPT-6.1 Terra, released October 26, 2026, as the cheaper option. OpenAI's estimate for Plus is 35 to 285 local messages per five hours, against 15 to 160 for GPT-6.1 Sol. These are estimates, and cloud tasks can use more.
For which GPT model suits which job, see which GPT model to use in 2026. For the two terminal-first agents side by side, Codex vs Claude Code has independent scores and cost per task. Teams building their own agents on the same machinery can read the Agents API pricing guide.
Windsurf, now Devin Desktop
Windsurf. As of this writing, windsurf.com 308-redirects straight to devin.ai/desktop. The product isn't "Windsurf, formerly known as" anything: the brand is retired. Documentation from the company behind it confirms existing users were migrated automatically, with settings, extensions and pricing carried over unchanged on the day of the switch.
What did change since: self-serve pricing moved off the old consumption model onto a flat ladder shared with Devin itself, the one in the cost list above.
Devin's self-serve billing docs, captured 19 Aug 2026. The ladder still matched this page when checked again on 24 Sep 2026; Devin Desktop and Devin's own cloud agent share it.
Devin, Tabnine and Trae
Devin. The fully autonomous end: give it a ticket, and it plans, writes, tests and opens a pull request without a human in the loop for most of the work. Legacy customers who were on the old per-Agent-Compute-Unit consumption plan were migrated to on-demand credits at "the same dollar value," per the company's own documentation, rather than losing their existing rate.
Tabnine. This is the entry that actually changed shape since August. Tricentis, an enterprise software-testing company, acquired Tabnine on July 30, 2026, to fold its context engine into Tricentis's own testing agents.
As of this run, tabnine.com no longer serves a coding-assistant pricing page at all: every path on the domain, including the root and /pricing, redirects to Tricentis's own site, landing specifically on a "contact us" page rather than a plan table. The self-serve tiers this guide cited in August are no longer independently verifiable, because there is no longer a live page quoting them. Treat any third-party site still quoting a self-serve Tabnine price as stale.
Tabnine is still real and still the closest thing here to a zero-retention, on-premises sale, but buying it now means starting a sales conversation with Tricentis, not signing up with a card.
Trae. The framing here also needs correcting. Trae's free tier is unchanged and still real: 5,000 autocompletions and two concurrent cloud tasks a month, no card required. But the paid ladder above it is the same three-price structure Cursor used until this change, with each tier carrying a monthly dollar usage allowance equal to its price rather than a completion count.
Trae's real differentiator now is the free tier, not a cheaper paid one: a team that can live inside 5,000 completions and two concurrent tasks a month still starts here for nothing, but a team that outgrows that free tier is no longer choosing Trae to save money over Cursor.
Data handling, security and compliance
Question four above named this the one most teams skip, right up until a security team exists to ask it. Three different answers sit on this shortlist:
- Cloud, vendor-hosted, engineer-reviewed. Cursor, Claude Code and Trae all process code on the vendor's own infrastructure by default, reviewed diff by diff by a human. Fine for most teams; a blocker the moment Legal wants a data-processing addendum the vendor hasn't signed yet.
- Cloud, but the agent can run on infrastructure the team already controls. Cursor's self-hosted machines (shipped September 2, 2026) keep a codebase, its build output and its secrets inside a company's own network while still using Cursor's cloud orchestration. It narrows the gap with Tabnine below without fully closing it.
- Sold specifically as a compliance product. Tabnine's entire pitch, pre-acquisition, was zero code retention and on-premises deployment as the default sale, not an enterprise add-on. That pitch still exists post-Tricentis, but it now starts with a sales call rather than a self-serve signup, which is itself a data point: a security review that used to start with a published SOC 2 report now starts with a meeting.
Devin and Devin Desktop's fully autonomous mode raises a related but different question: a human isn't reviewing every diff before it merges, so the data-handling review has to cover what the agent is trusted to do, not just what it can see.
The symlink flaw: who patched first
That trust has a dated test. On 8 July 2026 the security firm Wiz disclosed GhostApproval, a flaw it found in six coding assistants, three of them on this shortlist. A booby-trapped repository plants a symlink with an ordinary file name, the approval prompt shows that harmless name, and the agent's write lands on the real target, a developer's SSH key in the researchers' example.
Cursor had fixed it in version 3.0 on 5 June, weeks before the disclosure. Devin Desktop's fix came in version 3.6.1027 on 1 August 2026, 24 days after it: its edit and write tools now refuse to write through a symlink. The researchers tested the Windsurf IDE, not Devin's cloud agent. Anthropic called the scenario outside Claude Code's threat model and pointed to a symlink warning it added in February.
The hole is closed in both patched tools. What a security review can still weigh is which vendor moved before the researchers went public.
Cursor vs. Claude Code: the head-to-head most teams are actually having
These are the two tools most engineering teams are genuinely choosing between. The honest answer is that most professional teams that can afford it run both rather than picking one, because they solve different jobs from the framework above, not because either is incomplete on its own.
Where they actually diverge: Cursor gives you model choice inside a full editor with a free tier to start on, though third-party models now start at the $30 Premium plan; Claude Code has no free tier and lives in the terminal. Claude Code's 1-million-token context window is unchanged, but the benchmark most roundups cite for it is not: Anthropic has stopped reporting SWE-bench Verified for its current flagship and now publishes SWE-bench Pro instead, where Opus 5.5 scores 89.9%.
That is a different, harder benchmark introduced this cycle, not a direct upgrade on the old Verified number, so treat any comparison that lines the two scores up on one axis as sloppy rather than current. Hokai's model leaderboard tracks every GA model's current benchmark set and blended price side by side, which is the faster way to check whether a given score is still current than trusting a roundup's cached table.
If a team can only fund one, the surface matters more than either benchmark: pick the editor tool if the team's daily work happens in an IDE, pick the terminal tool if it happens in CI pipelines and long-running repo tasks.
Where GitHub Copilot fits in
Notably absent from either company's own marketing: GitHub Copilot, still the tool most developers meet first. Its model access has widened: Free includes access to Claude Haiku 4.5 and GPT-5 mini, and Pro bundles third-party agents, specifically Claude Code and Codex, inside the same $10/month seat.
Claude Opus 5.5 became selectable inside Copilot on September 22, 2026, the same day it became the Claude Code default. GitHub's plans page now lists Opus under the premium models of Pro+, so the cheaper-than-Claude-Pro route that existed at launch should be checked against the plan page before a team relies on it.
The newer change is metering. Every paid plan carries monthly AI credits, from $15 on Pro up to $200 on Max, and agent mode, code review, cloud agents and the command line all consume them.
GitHub Copilot's plans page, captured 19 Aug 2026. The four prices were the same when the page was reread on 6 Nov 2026; what each plan includes has since expanded to monthly AI credits.
Not on this shortlist, on purpose
Sourcegraph, Greptile and Lightrun all sit in Hokai's coding-assistant category, and all three get left off "best coding assistant" shortlists for a bad reason: they don't write code, so they lose a head-to-head they were never entered in.
Sourcegraph's free and individual Pro plans were discontinued on July 23, 2025; what's left is an Enterprise product starting around $16,000, built for searching and navigating codebases at a scale where "which assistant writes the best function" stops being the question.
Greptile does one job, automated pull-request review. Its Starter plan is free for one active developer with 50 credits a month, and Pro has no seat fees: unlimited developers and agents, with credits bought at $1 each. It is a genuine complement to any tool on the shortlist above, not a competitor to it. Lightrun does the opposite job: it lets a team inspect variables, logs and traces in a running production service without redeploying, which has nothing to do with generating code in the first place.
Two open-source terminal agents also sit outside the shortlist. Cline is free, runs in VS Code, JetBrains and the terminal, and has you bring your own model key; its optional ClinePass add-on is $9.99/month. Gemini CLI is Google's open-source terminal agent, but after Google's June 2026 consumer shutdown it is limited to enterprise Gemini Code Assist licences, so it is no longer a free personal option.
Google Antigravity is the newest name a team will be asked about. Google sells it as an app, a command-line tool, an editor and an SDK, with a free individual plan and higher limits through its AI plans.
Hokai has not tested it and has no record for it yet, so it stays off the shortlist: no source read this run puts it ahead of the picks above on any job. It is the one to trial if a team already pays Google for AI.
A team evaluating "the best AI coding assistant" for Sourcegraph, Greptile or Lightrun is solving the wrong problem; the right question is whether the team's actual bottleneck is writing code at all.
Who should skip this category entirely
Not every team with developers needs a pick from this list. A solo founder writing a few hundred lines a week gets more value from a free ChatGPT or Claude chat tab than from a $20/month subscription to a tool built around repo-wide agentic workflows they'll never trigger.
A team whose real bottleneck is process, not typing speed, should fix that first: adding Cursor or Claude Code on top of a broken code-review or deploy pipeline speeds up the part that was never the constraint. And a team already deep in one vendor's ecosystem with a signed security review, GitHub Copilot inside an all-Microsoft shop, for instance, usually gets more from pushing that tool harder than from adding a second one from this list for a marginal capability gain.
The same job-first logic applies one level down. A platform engineer's shortlist looks different from an app developer's, which is why Hokai runs a separate guide to the best AI tools for DevOps and platform engineers rather than folding that job into this one.
A team that needs everything running on hardware it owns, with no cloud call at all, is answering a different question again, covered in the best AI models for local coding. And a developer who just wants quick answers to coding questions without adopting an IDE agent has a lighter option covered in the free AI tools for coding questions.
The turn: procurement doesn't shop by job
The framework above assumes a team is free to pick the tool that fits the job. Above a certain company size, that stops being true: once Legal has cleared one vendor's SOC 2 report and data-handling terms, buying a second, differently-shaped product from that same vendor is faster than restarting a job-based evaluation from scratch.
That's the more cynical read on why Windsurf's billing was folded into Devin's rather than launching a fifth, job-specific product: consolidation under one procurement approval beats specialization once security review gets involved.
For a team under about ten engineers with no procurement process to route around, the job-based framework above still holds: there's no vendor-lock advantage to capture when nobody's cleared anybody yet. Above that size, expect the shortlist to compress toward whichever vendor's paperwork is already signed, regardless of which job actually needs solving.
Tabnine's acquisition by Tricentis is a different version of the same forcing function, one layer up: it wasn't a coding-assistant company buying scale, it was a testing-and-quality company buying a context engine. The vendor on the other side of a "coding assistant" purchase may not be optimizing for coding assistants at all a year from now.
What to watch
Google shipped Gemini 3.7 Flash on August 13, 2026, a model built specifically for coding and agentic workflows, at an introductory $0.75-per-million-input-token rate through the end of 2026. Six weeks on, it's still available only in Google's own tools (Android Studio, AI Studio, Gemini Enterprise), not as a selectable model inside any assistant on this shortlist.
If anything, the model layer moved around it instead: Claude Opus 5.5 is now the default in two of the tools above, and it took less than a day for the second one to add it after the first. The moment a Cursor, Devin or Trae adds Gemini 3.7 Flash to its own model pool, model choice, not vendor choice, becomes the next axis this category gets ranked on; as of this run, that moment hasn't arrived.
The frontier isn't only Anthropic, OpenAI and Google anymore, either. On September 2, 2026, Alibaba's Qwen3.8-Max refresh took the top spot on Code Arena WebDev, ahead of Claude Opus 5. Eight days later, DeepSeek shipped DeepSeek V4.1 Flash, an open-weight model that VentureBeat reported beating GPT-6 Sol and the previous Claude Opus 5 on agentic-coding benchmarks, at a fraction of the per-token price.
None of that has reached this shortlist's tools yet. A team already choosing Cursor's self-hosted machines or Tabnine's compliance pitch for control-over-infrastructure reasons should know a fully open model layer, not just an open agent, is starting to exist too.
Google's Gemini 3.7 Flash announcement, captured 19 Aug 2026. As of 24 Sep 2026, none of the assistants on this shortlist list it as a selectable model yet.
If your shortlist is down to the two terminal-first agents, read Codex vs Claude Code for independent scores, cost per task and plan limits.
Frequently asked questions
What's the best AI coding assistant for a small startup team?
For most 4-20 person teams, the honest starting stack is one IDE-native tool for daily writing (Cursor's free plan, or Premium at $30/month once third-party models are needed) plus Claude Code ($20/month) for terminal-driven, repo-wide work. Trae's free tier (5,000 completions a month, no credit card) is still the real budget option; its paid ladder of $20, $60 and $200 a month copies Cursor's old pricing.
Is Windsurf still a separate product from Devin?
No. Cognition retired the Windsurf brand on June 2, 2026 and folded it into Devin Desktop; windsurf.com now redirects directly to devin.ai/desktop. Existing users were migrated automatically, and Devin Desktop shares the same Free / $20 Pro / $200 Max / $80-minimum Teams pricing ladder as Devin's own autonomous agent.
How much does Claude Code cost, and is there a free tier?
Claude Code has no standalone free tier. It ships inside Anthropic's paid Claude plans, starting at Pro for $20/month ($17/month billed annually), then Max at $100/month and Ultra at $200/month for heavier use. On the Anthropic API its model aliases now resolve to Opus 5.5, Sonnet 5.5 and Haiku 5.5, and Fable 5.1 is opt-in and can bill to usage credits on some plans.
Are Sourcegraph, Greptile and Lightrun AI coding assistants?
Not in the code-writing sense. Sourcegraph is enterprise codebase search (from roughly $16,000, after its free and individual plans were discontinued in July 2025), Greptile is automated pull-request review (free for one developer, $30/seat/month for teams), and Lightrun is live production debugging. All three complement a coding assistant rather than replacing one.
What happened to Tabnine?
Tricentis, an enterprise software-testing company, acquired Tabnine on July 30, 2026, to use its context engine inside Tricentis's own testing agents. As a result, tabnine.com no longer publishes a self-serve pricing page: every path, including /pricing, now redirects to Tricentis's own site and a 'contact us' form. Tabnine still exists as a private, compliance-focused coding assistant, but buying it now starts with an enterprise sales conversation rather than a credit card.
Covered in this guide
- Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
- Anthropic: Anthropic is a Public Benefit Corporation started by 7 ex-OpenAI researchers. It builds the Claude model family and counts Amazon and Google among its largest backers.
- ChatGPT: ChatGPT is OpenAI's AI assistant, with 1.2 billion weekly users by OpenAI's count, the GPT-6 model family and plans that run from a free tier to a premium Pro tier.
- Claude Haiku 5.5: Anthropic's smallest and fastest 5.5 model, released October 7, 2026, with a 1M context window, built for high-volume work.
- Claude Opus 5.5: Anthropic's September 2026 flagship LLM with a 1 million token context window, built for long-running agentic coding and computer use.
- Claude Sonnet 5.5: Anthropic's fastest Sonnet-class model, launched in September 2026, with a 1M context and output more than 30% faster than Sonnet 5.
- Cline: Free, open source AI coding agent for VS Code, JetBrains, and the terminal with 11M+ installs; bring your own Claude, GPT, or Gemini key.
- Codex Security: OpenAI's application security agent scans connected GitHub repositories, tests likely bugs in a sandbox, and drafts fixes for review as pull requests.
- DeepSeek: DeepSeek is a Chinese AI research company developing frontier language and reasoning models, including DeepSeek V4-Pro. Founded by High-Flyer hedge fund CEO Liang Wenfeng in 2023, the company is known for achieving GPT-level performance at dramatically lower compute and API cost.
- DeepSeek V4.1 Flash: The MIT-licensed vision-and-text model, launched in September 2026 and refreshed with an October snapshot.
- Gemini 3.7 Flash: Google DeepMind's August 2026 Flash model for coding and agents with a 1M-token context window; Google now recommends a newer Flash model and routes requests to it.
- Gemini CLI: Google's open-source coding agent for the terminal, now limited to enterprise Code Assist licenses after its June 2026 consumer shutdown.
- GitHub: Founded in 2008, GitHub is the largest code repository and collaboration platform with 140M+ developers, now integrated into Microsoft's CoreAI division after CEO Thomas Dohmke departed end of 2025.
- Copilot: GitHub's AI coding assistant with agent mode, Copilot Fusion model routing and MCP support across 6 IDEs, plus a free plan with no credit card required.
- GPT-6 Luna: GPT-6 Luna (Sept 2026) is OpenAI's lowest-cost GPT-6 model, tuned for high-volume chat, extraction and classification with six adjustable reasoning-effort levels.
- GPT-6 Sol: OpenAI's mid-tier GPT-6 model (Sep 2026), 1.05M-token context, 68.8% on DeepSWE v1.1 coding, half GPT-5.6 Sol's price.
- GPT-6.1 Sol: GPT-6.1 Sol is OpenAI's September 2026 mid-tier reasoning model for coding, computer use and document work, built to sit just below GPT-6 Astra.
- GPT-6.1 Terra: GPT-6.1 Terra is OpenAI's October 2026 model for computer use, image reading and writing, built for high-volume agent work.
- Lightrun: AI SRE platform that adds live logs, snapshots and metrics to running production apps without redeployment, now with autonomous incident detection and remediation.
- OpenAI Codex: Clones a GitHub repo into a sandbox, writes and tests code, then opens a pull request; included in every ChatGPT plan, with the newest GPT-6 models from Plus.
- Qwen3.8-Max: Qwen3.8-Max is Alibaba's flagship MoE model, launched Aug 2026, with a 1M-token context, native multimodal input, and a text-only open-weight checkpoint released weeks later.
- Sourcegraph: Enterprise code search and AI Deep Search for multi-repo codebases. Its Cody assistant is being retired in January 2027.
- Tabnine: Enterprise AI coding assistant with privacy-first architecture, agentic workflows, and flexible deployment
- Trae: Free AI IDE by ByteDance with SOLO autonomous agent and 4 plans from $3/month; supports Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and DeepSeek on macOS, Windows, Web, and iOS.
- Windsurf: Windsurf, now Devin Desktop from Cognition, is a VS Code-based editor with local and cloud agents, free with unlimited Tab and Pro at $20 a month.
Sources
Still deciding?
This guide covers a handful of options. Smart Match checks every listing in the directory against how you actually work and what you can spend, then hands you the shortlist and the reason behind each pick.
Find your tools

