AI coding assistants have just been through their most disruptive year yet. Between April and June 2026, every major vendor moved to metered credit or token billing: OpenAI switched Codex to token-based consumption from 2 April, GitHub replaced Copilot’s premium requests with AI Credits on 1 June, Windsurf disappeared into Devin Desktop on 2 June, and Google gutted the Gemini CLI free tier on 18 June. If your mental map of this market predates that ten-week upheaval, it is out of date.
This guide compares the five tools most teams shortlist — Claude Code, GitHub Copilot, Cursor, OpenAI Codex and Devin Desktop — plus the wider field of Google Antigravity 2.0, AWS Kiro and the open-source agents. It lives inside our AI models and tools hub, where you will also find model-level explainers that this comparison deliberately does not repeat. Every price here comes from vendors’ own published pages, including GitHub’s official Copilot plan listing, and reflects the market as of 14 August 2026.
We compare AI coding assistants on three axes rather than a feature checklist: the workflow surface each tool occupies, what a metered dollar actually buys on each plan, and the enterprise controls behind the marketing pages. If you want working habits rather than tool selection, our guide to AI coding best practices covers that ground; this piece stays firmly on tools and pricing.
One naming correction before we start, because half the round-ups still get it wrong. Windsurf no longer exists as a brand: Cognition renamed it Devin Desktop on 2 June 2026, and windsurf.com now permanently redirects to devin.ai. And OpenAI’s product is simply Codex — a CLI, an IDE extension and cloud tasks bundled into ChatGPT plans. Getting the names right is the first step to comparing the best AI coding assistants on their 2026 merits.
Table of contents
- Why AI Coding Assistants Look Different in 2026
- The Three Workflow Surfaces of AI Coding Assistants
- Claude Code: The Terminal Agent to Beat
- GitHub Copilot: The Incumbent Rebuilt Around AI Credits
- Cursor: The IDE That Built Its Own Frontier Model
- OpenAI Codex: Bundled Into ChatGPT
- Devin Desktop: What Happened to Windsurf
- Pricing Compared: What AI Coding Assistants Cost in 2026
- Agentic Features Compared Across AI Coding Assistants
- Enterprise Controls in AI Coding Assistants
- The Wider Field: Antigravity, Kiro and Open-Source Agents
- How to Choose Between AI Coding Assistants in 2026
- Frequently Asked Questions
- References
Why AI Coding Assistants Look Different in 2026
The story of 2026 so far is not a new model or a new editor. It is billing. Within roughly ten weeks, the economics of AI coding assistants were rebuilt around metered consumption, and the market collapsed into a $20 entry tier and a set of $100–$200 power tiers. Understanding that shift explains almost every pricing table in this guide.
The Ten-Week Repricing That Hit AI Coding Assistants
Four dates define the shift. From 2 April 2026, Codex usage on ChatGPT Plus, Pro, Business and Enterprise began billing against credit and token consumption rather than task counts. On 1 June, GitHub replaced premium-request counting with usage-based GitHub AI Credits. On 2 June, Cognition rebranded Windsurf as Devin Desktop via an over-the-air update. And on 18 June, Google stopped serving Gemini CLI requests for free, AI Pro, AI Ultra and individual Code Assist accounts, cutting the free ceiling from roughly 1,000 requests per day to the low tens.
Any comparison of AI coding assistants written before April 2026 is therefore describing a different market. The tools kept their names (mostly), but the meters behind them changed completely.
From Request Counts to Metered Tokens
The new billing pattern is remarkably consistent. GitHub’s AI Credits peg 1 credit to $0.01, with usage metered on input, output and cached tokens at each model’s published rate. OpenAI’s own estimate puts typical Codex power-user spend at $100–$200 per developer per month under the new token billing. AWS Kiro sells add-on credits at $0.04 each, and Google Antigravity tops up 2,500 credits for $25.
The practical consequence: the sticker price of AI coding assistants now tells you less than the metered allowance behind it. A $10 plan with $15 of credits and a $20 plan with a generous compute pool can cost the same in real use — or differ by a factor of five.
A Market That Collapsed Into Two Price Points
Line the plans up and a pattern emerges. Entry pricing has converged hard on $20 per month: Claude Pro at $17–$20, Cursor Pro at $20, ChatGPT Plus at $20, Devin Pro at $20, Google AI Pro at $20 and Kiro Pro at $20. Above that sits a power band: Claude Max from $100, Copilot Max at $100, then Cursor Ultra, Devin Max, ChatGPT Pro’s 20x tier and Kiro Power all at $200.
Industry tracking of AI coding assistants puts real heavy-usage costs at $60–$200 per month, typically split across two tools. We return to that two-tool pattern in the buying advice at the end.
The Three Workflow Surfaces of AI Coding Assistants
Feature lists make AI coding assistants look interchangeable. Workflow surfaces do not. In 2026 the market splits cleanly into three shapes — IDE-native, terminal-first and agent-manager — and the right choice depends far more on which surface fits your team than on any benchmark.
IDE-Native AI Coding Assistants
GitHub Copilot and Cursor anchor this category: the assistant lives inside the editor, completing lines, chatting in a sidebar and increasingly running agents against your open workspace. Copilot reaches the largest installed base through VS Code and github.com, while Cursor ships an entire forked IDE built around its agent. IDE-native AI coding assistants remain the lowest-friction starting point, because nobody has to change how they work on day one.
Terminal-First AI Coding Assistants
Claude Code and the Codex CLI treat the terminal, not the editor, as home. That suits long-running agentic work: delegating a refactor, running tests, spawning subagents. GitHub joined this camp when Copilot CLI went generally available in February 2026 with agent capabilities of its own. Terminal-first AI coding assistants tend to be the “second tool” in the pairing pattern — the heavy-lifting complement to an IDE assistant.
Agent-Manager Platforms
The newest surface inverts the IDE entirely. Cognition positions Devin Desktop as “a full IDE with an agent manager built in, not the other way around”, with its Agent Command Center replacing the editor canvas as the default view. Google Antigravity 2.0 bundles a desktop IDE, CLI, SDK and a Managed Agents tier in the Gemini API. AWS Kiro organises everything around specs, subagents and hooks. Here you supervise a fleet of agents rather than drive a single one.
Claude Code: The Terminal Agent to Beat
Anthropic’s Claude Code is the reference point most 2026 comparisons of AI coding assistants measure against, and the benchmark numbers explain why. It is a terminal agent first, with IDE and web surfaces layered on top, and it inherits whatever frontier model Anthropic ships next.
Claude Code Pricing Tiers in 2026
Claude Code is included in Claude Pro at $17 per month on annual billing ($20 monthly), which also bundles Claude Cowork and access across web, desktop and terminal. Claude Max starts at $100 per month with a choice of 5x or 20x Pro usage, higher output limits and priority access at peak times.
For organisations, Claude Team costs $20 per seat per month annual ($25 monthly) for 2–150 members with admin controls, SSO and audit capabilities — but note the catch: Claude Code is only included on Premium seats at $100 per seat per month annual ($125 monthly), with a five-seat minimum. Budget accordingly.
Subagents, Nested Spawning and Agent Teams
Claude Code’s headline 2026 agentic features are subagents with isolated context windows and their own tool loadouts, nested agent spawning capped at depth 5, and experimental Agent Teams enabled via the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 flag. In practice this means a single session can decompose work — one subagent runs tests while another edits, each with a clean context.
Among AI coding assistants, this is currently the deepest first-party multi-agent stack; most rivals either run one agent at a time or push orchestration to a cloud service.
The Models Behind Claude Code
Anthropic’s 2026 cadence has been relentless: Opus 4.6 on 5 February, Sonnet 4.6 on 17 February, Opus 4.7 on 16 April, Opus 4.8 on 28 May, Mythos 5 and Fable 5 on 9 June, Sonnet 5 on 30 June and Opus 5 on 24 July. Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8 — and is now the default model inside Claude Code.
As of 5 August 2026, Claude Fable 5 leads agentic coding at 95.0% on SWE-bench Verified and 80.3% on SWE-bench Pro. No rival family matches those figures, which is why so many competing AI coding assistants also resell Claude models.
GitHub Copilot: The Incumbent Rebuilt Around AI Credits
Copilot remains the default answer when developers first try AI coding assistants, but the product under the name changed substantially in 2026. It is now a credit-metered platform that hosts other vendors’ agents — including Claude Code and Codex — inside its own experience.
Copilot’s Free, Pro, Pro+ and Max Tiers
The August 2026 individual line-up runs four deep. Free costs nothing and includes 2,000 completions and 50 chat or agent requests monthly, powered by Claude Haiku 4.5 and GPT-5 mini. Pro costs $10 per month and carries $15 in GitHub AI Credits. Pro+ costs $39 with $70 in credits and Claude Opus access. The new Max tier costs $100 per month with $200 in credits, aimed at sustained agent workflows.
That Free tier remains the cheapest legitimate way to trial AI coding assistants at all — a genuine $0 entry the rest of the market has largely abandoned.
How GitHub AI Credits Work
Since 1 June 2026, every Copilot plan meters usage in GitHub AI Credits: 1 credit equals $0.01, and consumption is calculated on input, output and cached tokens at each model’s published rate. Choosing an expensive model burns the same allowance faster, so the effective monthly capacity of a Pro plan depends heavily on whether you run a frontier model or a budget one.
Third-Party Agents and the Model Picker
Copilot’s most strategic 2026 move was opening up. Copilot Pro includes third-party agent integration, letting Claude Code and Codex run inside the Copilot experience alongside GitHub’s own cloud agents and code review. Since 14 April 2026, github.com supports model selection for both Claude and Codex agents, and the 2026 model picker spans Claude Opus 4.8, Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5.5, GPT-5.4, GPT-5.3-Codex and Gemini 3.1 Pro. Copilot is quietly becoming a marketplace for rival AI coding assistants rather than a single assistant.
Copilot Business and Enterprise
Organisational pricing sits on top of the individual tiers: Copilot Business costs $19 per user per month and Copilot Enterprise $39 per user per month. Both inherit the AI Credits metering, which makes cost auditing a new line item for platform teams.
Cursor: The IDE That Built Its Own Frontier Model
Cursor took the opposite bet from every other maker of AI coding assistants: rather than reselling frontier models alone, Anysphere trained its own. That vertical integration now defines both its performance story and its pricing.
Cursor Pricing From Hobby to Ultra
Cursor’s individual tiers as of August 2026: Hobby is free, Pro costs $20 per month, Pro+ costs $60 with 3x Pro’s agent limits, and Ultra costs $200 with 20x. Annual billing takes 20% off every paid plan — the most generous annual discount among the major AI coding assistants. Teams costs $40 per user per month, and Enterprise pricing is custom.
Composer 2 and Composer 2.5
Composer 2, Anysphere’s in-house frontier coding model released 19 March 2026, pairs a 200K context window with a mixture-of-experts architecture, costs $0.50 per million input tokens and $2.50 per million output tokens, and is exclusive to the Cursor IDE — there is no external API. Composer 2.5 followed on 18 May 2026, tuned for long agentic sessions, with benchmark results rivalling Claude Opus 4.7 on SWE-Bench and Terminal-Bench. A cheap, fast in-house model is Cursor’s answer to the metered-billing squeeze.
What Cursor Pro Actually Includes
The $20 Pro plan is unusually complete: extended agent limits, frontier model access, MCPs, skills, hooks, cloud agents and usage-based Bugbot for automated bug-hunting. Where other AI coding assistants gate agentic features behind power tiers, Cursor’s differentiators mostly arrive at entry level, with the higher tiers selling capacity rather than capability.
OpenAI Codex: Bundled Into ChatGPT
Codex is the only one of the big five you cannot buy on its own. It ships inside ChatGPT plans as a CLI, an IDE extension and cloud tasks — which makes it either a free bonus or an awkward procurement line, depending on whether your organisation already pays for ChatGPT.
Codex Pricing Through ChatGPT Plans
The August 2026 ladder: ChatGPT Free offers limited Codex access, Go costs $8 per month, Plus $20, and Pro $100 per month for 5x limits or $200 for 20x. Business seats cost $25 per user per month ($20 annual) and Enterprise is custom. Since 2 April 2026 all of it bills against credit and token consumption rather than task counts, and OpenAI’s own estimate puts typical power-user spend at $100–$200 per developer per month — worth quoting to anyone who still thinks $20 buys unlimited agentic coding from any of the AI coding assistants.
GPT-5.6 Sol, Terra and Luna
Codex’s recommended model family as of August 2026 is GPT-5.6: flagship Sol as the default at medium reasoning effort, balanced Terra and cost-efficient Luna, plus the ChatGPT Pro-only GPT-5.3 Codex Spark real-time preview. GPT-5.6 reached general availability on 9 July 2026 after a preview that opened 26 June; Sol is priced at $5 per million input tokens and $30 per million output tokens, identical to the previous GPT-5.5 flagship. Note the housekeeping deadline: GPT-5.4 and GPT-5.4 mini retire from Codex on 31 August 2026.
What Codex Shipped in August 2026
Codex’s August 2026 releases added experimental paginated thread history with sub-agent support and memories, audio inputs, and an expanded /import command that migrates Cursor and Claude Code settings, MCP servers, plugins, sessions and project-scoped memories. That /import is a pointed feature: OpenAI clearly expects developers to be switching between AI coding assistants, and wants the move to take minutes.
Devin Desktop: What Happened to Windsurf
If you last evaluated Windsurf in 2025, the product you remember now answers to a different name and a different philosophy. Of all the AI coding assistants in this guide, this is the one whose 2026 story is most often reported wrongly.
From Windsurf to Devin Desktop
The timeline: Cognition acquired Windsurf — product, IP, trademark and team — on 14 July 2025, then rebranded it Devin Desktop on 2 June 2026 via an over-the-air update, with plans, pricing, settings and extensions carrying over unchanged. windsurf.com now permanently redirects (HTTP 308) to devin.ai. Existing users kept everything; the brand simply vanished.
Devin Desktop Pricing in August 2026
Devin’s August 2026 pricing: Free at $0, Pro at $20 per month with OpenAI, Claude and Gemini models plus free access to Cognition’s own SWE 1.7, Max at $200 per month, Teams at $80 per month base plus $40 per developer seat, and custom Enterprise with SAML/OIDC SSO and dedicated deployment. The Teams structure — a platform fee plus per-seat pricing — is unique among the mainstream AI coding assistants and rewards larger teams.
The Agent Command Center
Cognition’s positioning line is the clearest statement of the agent-manager thesis: Devin Desktop is “a full IDE with an agent manager built in, not the other way around”. The Agent Command Center replaces the editor canvas as the default view — you open the tool to a fleet of agents and their tasks, and drop into code only when needed. Whether that inversion suits you is the single most important question in evaluating it.
Pricing Compared: What AI Coding Assistants Cost in 2026
Put side by side, the pricing of AI coding assistants in August 2026 shows the convergence described earlier: a crowded $20 entry shelf and a power band from $100 to $200. The table below gathers every published individual and team tier from the vendors’ own pages.
| Tool | Free tier | Entry paid tier | Power tiers | Team / Business |
|---|---|---|---|---|
| Claude Code | — | Pro $17/mo annual ($20 monthly) | Max from $100/mo (5x or 20x Pro) | Team $20–$25/seat; Claude Code on Premium seats $100/seat annual, min 5 |
| GitHub Copilot | $0: 2,000 completions, 50 chat/agent requests | Pro $10/mo + $15 credits | Pro+ $39 + $70 credits; Max $100 + $200 credits | Business $19/user; Enterprise $39/user |
| Cursor | Hobby | Pro $20/mo | Pro+ $60 (3x); Ultra $200 (20x); annual −20% | Teams $40/user; Enterprise custom |
| OpenAI Codex | ChatGPT Free (limited) | Go $8; Plus $20 | Pro $100 (5x) or $200 (20x) | Business $25/user ($20 annual); Enterprise custom |
| Devin Desktop | Free $0 | Pro $20/mo | Max $200/mo | Teams $80 base + $40/seat; Enterprise custom |
| Google Antigravity | Individual: unlimited Tab and Command | Google AI Pro $20/mo | AI Ultra $99.99/mo (5x) or $199.99/mo (20x) | Consumption-based via Google Cloud |
| AWS Kiro | Free: 50 credits/mo | Pro $20 (1,000 credits) | Pro+ $40; Pro Max $100; Power $200 (10,000) | GovCloud roughly 20% higher |
The $20 Entry Tier Across AI Coding Assistants
Six vendors now sell an entry plan at effectively the same price: Claude Pro ($17 annual, $20 monthly), Cursor Pro ($20), ChatGPT Plus with Codex ($20), Devin Pro ($20), Google AI Pro ($20) and Kiro Pro ($20). GitHub undercuts them all at $10 with $15 of credits, and ChatGPT Go sits at $8 for lighter use.
At this level the differentiator is not price but what the allowance covers. Cursor’s $20 includes its in-house Composer models at $0.50/$2.50 per million tokens — cheap capacity the metered rivals cannot match. Claude Pro’s $17–$20 includes Claude Code plus Claude Cowork. Kiro’s $20 buys a defined 1,000 credits. Comparing AI coding assistants at the entry tier means comparing allowances, not stickers.
The Power Tiers of AI Coding Assistants
The takeaway before the chart: five flagship individual tiers now cluster at the $200 mark, with Google’s AI Ultra topping out at $199.99 per month for its 20x tier since the May 2026 I/O restructure retired the old $249.99 plan, while Claude Max and Copilot Max open the power band at $100. All figures below are the vendors’ published monthly prices from the table above.
The $200 cluster is no coincidence: it matches OpenAI’s $100–$200 estimate of real power-user consumption. Vendors have priced the top of the market at what heavy agentic use actually costs to serve.
Hidden Multipliers: Credits, Pools and Top-Ups
The fine print moves real-world costs more than the headline tiers. GitHub meters at 1 credit = $0.01 against each model’s token rate. Kiro sells add-on credits at $0.04 each and charges roughly 20% more in GovCloud. Antigravity’s usage pool is compute-based and refreshes every 5 hours, with top-ups at 2,500 credits for $25. When budgeting for AI coding assistants, model choice inside the tool is now a bigger cost lever than plan choice.
Agentic Features Compared Across AI Coding Assistants
Autocomplete is a solved problem; the 2026 competition between AI coding assistants is about agents — how many can run at once, how they are supervised, and which models drive them. The table summarises the published agentic capabilities of each platform.
| Tool | Workflow surface | Multi-agent capability | Default / flagship models |
|---|---|---|---|
| Claude Code | Terminal agent | Subagents with isolated contexts, nested spawning to depth 5, experimental Agent Teams | Claude Opus 5 (default) |
| GitHub Copilot | IDE-native + CLI (GA Feb 2026) | Cloud agents, code review, third-party agents (Claude Code, Codex) | Picker: Opus 4.8, Sonnet 4.6, Haiku 4.5, GPT-5.5, GPT-5.4, GPT-5.3-Codex, Gemini 3.1 Pro |
| Cursor | IDE-native | Cloud agents, hooks, skills, MCPs, usage-based Bugbot | Composer 2 / 2.5 plus frontier model access |
| OpenAI Codex | CLI, IDE extension, cloud tasks | Experimental sub-agent thread history, memories, /import migration | GPT-5.6 Sol (default), Terra, Luna; Codex Spark preview on Pro |
| Devin Desktop | Agent-manager IDE | Agent Command Center as default view | SWE 1.7 free, plus OpenAI, Claude and Gemini models |
| Antigravity 2.0 | IDE, CLI, SDK | Managed Agents tier in the Gemini API | Gemini 3.5 Flash (default), Gemini 3.1 Pro, Claude Sonnet and Opus 4.6, gpt-oss-120b |
| AWS Kiro | Spec-driven IDE + CLI | Specs, custom subagents, powers, hooks | Frontier Anthropic model; Qwen3 Coder Next and DeepSeek v3.2 on Free |
Multi-Agent Orchestration in AI Coding Assistants
Claude Code offers the deepest first-party stack: subagents with isolated context windows and tool loadouts, nesting to depth 5, and experimental Agent Teams behind an environment flag. Codex is moving the same way, with August 2026’s experimental paginated thread history adding sub-agent support and memories. Devin Desktop and Antigravity take a management view — fleets of agents supervised from a command centre or a Managed Agents API tier — while Kiro binds its subagents to written specs. The orchestration gap between AI coding assistants is now wider than the model gap.
Model Choice Inside AI Coding Assistants
Openness varies sharply. Copilot’s picker spans seven models across three vendors, and github.com has supported model selection for Claude and Codex agents since 14 April 2026. Devin Pro bundles OpenAI, Claude and Gemini models plus its own SWE 1.7; Antigravity’s free tier spans Gemini, Claude and gpt-oss-120b; Kiro pairs a frontier Anthropic model with open-weight Qwen3 Coder Next and DeepSeek v3.2. At the closed end, Cursor’s Composer 2 is exclusive to its IDE with no external API, and Claude Code runs Claude models only. Multi-vendor teams should weigh how much flexibility their AI coding assistants give them here.
Benchmarks: SWE-bench Verified and SWE-bench Pro
Two published August 2026 datapoints frame the model race. Morph’s tracking has Claude Fable 5 leading agentic coding at 95.0% on SWE-bench Verified and 80.3% on SWE-bench Pro. And OpenAI’s own July 2026 comparison shows GPT-5.6 Sol at 64.6% on SWE-Bench Pro against Claude Mythos 5’s 80.3%, with Terra at 63.4% and Luna at 62.7%. The chart plots those published SWE-bench Pro scores.
Treat benchmarks as one signal among several: Composer 2.5’s published results rival Claude Opus 4.7 on SWE-Bench and Terminal-Bench, and real-world fit depends on your codebase, not a leaderboard.
Enterprise Controls in AI Coding Assistants
For a ten-person startup, plan price decides. For a regulated enterprise, the governance layer decides — and here the AI coding assistants diverge far more than their pricing pages suggest. The table collects each vendor’s published organisational controls.
| Tool | Team pricing | Enterprise | Published controls |
|---|---|---|---|
| Claude / Claude Code | Team $20–$25/seat (2–150); Premium seats $100–$125, min 5 | Custom | Managed settings hierarchy, managed-settings.d/ fragments, server-managed policies, OS-level sandbox, 500K context, SSO, SCIM, compliance API, HIPAA-ready |
| GitHub Copilot | Business $19/user | Enterprise $39/user | AI Credits metered at published per-model token rates |
| Cursor | Teams $40/user | Custom | SAML/OIDC SSO, team-wide privacy mode, agentic code reviews, usage analytics; Enterprise adds pooled usage, SCIM, access controls, audit logs |
| OpenAI Codex | ChatGPT Business $25/user ($20 annual) | Custom | Token-consumption billing across Business and Enterprise since 2 April 2026 |
| Devin Desktop | Teams $80 base + $40/seat | Custom | SAML/OIDC SSO, dedicated deployment |
| Antigravity 2.0 | Via Google Cloud | Enterprise path | Consumption-based organisation pricing through Google Cloud |
| AWS Kiro | Power $200/user | — | GovCloud availability at roughly 20% price premium |
Policy Enforcement and Sandboxing
Claude Code publishes the most detailed control stack of the AI coding assistants compared here. Managed settings sit at the top of a policy hierarchy that nothing can override, a managed-settings.d/ directory merges policy fragments, and server-managed policies can be pushed from Anthropic’s admin console. Beneath that, an OS-level sandbox — macOS Seatbelt, Linux seccomp and namespaces — enforces filesystem and network boundaries on Bash commands the agent runs. For security teams nervous about agents executing shell commands, that sandbox is the headline feature.
Identity, SSO and SCIM
Identity support is now table stakes but arrives at different tiers. Claude Team includes SSO and audit capabilities from $20–$25 per seat, with SCIM, a compliance API, HIPAA-ready configuration and a 500K context window on Claude Enterprise. Cursor puts SAML/OIDC SSO on Teams at $40 per user and SCIM on Enterprise. Devin offers SAML/OIDC SSO with dedicated deployment on its custom Enterprise plan. Check which tier your identity requirements actually land on before comparing prices of AI coding assistants — it routinely doubles the real per-seat cost.
Seat Pricing for Teams
The seat maths favours different tools at different sizes. Copilot Business at $19 per user is the cheapest managed entry. Cursor Teams at $40 sits mid-pack. Devin’s $80 base plus $40 per seat penalises tiny teams but converges towards $40 as headcount grows. Claude is the outlier: a general Team seat costs $20–$25, but a Premium seat with Claude Code costs $100–$125 with a five-seat minimum, so a mixed deployment needs careful role mapping. An AI strategy engagement typically starts exactly here — matching seat types to who genuinely needs agentic capability.
The Wider Field: Antigravity, Kiro and Open-Source Agents
The big five are not the whole market. Google and AWS both re-entered the fight in 2026, and open-source AI coding assistants now carry six-figure GitHub star counts. Any shortlist should at least glance at this field.
Google Antigravity 2.0 and the Gemini CLI Squeeze
Antigravity 2.0, launched at I/O 2026, bundles a desktop IDE, CLI, SDK, a Managed Agents tier in the Gemini API and an enterprise path, defaulting to Gemini 3.5 Flash with a compute-based usage pool that refreshes every 5 hours. The free Individual plan is aggressive: unlimited Tab completions and Command requests, with models spanning Gemini 3.5 Flash, Gemini 3.1 Pro, Gemini 3 Flash, Claude Sonnet and Opus 4.6 and gpt-oss-120b.
The generosity has a flip side. On 18 June 2026, Google cut the Gemini CLI free ceiling from roughly 1,000 requests per day to the low tens, steering individual users towards the Antigravity CLI or a paid Gemini API key. Free tiers among AI coding assistants can evaporate mid-subscription — plan for it.
AWS Kiro’s Credit Ladder
Kiro’s 2026 tiers climb evenly: Free with 50 credits per month, Pro at $20 (1,000 credits), Pro+ at $40 (2,000), Pro Max at $100 (5,000) and Power at $200 per user per month (10,000), with add-on credits at $0.04 each and roughly 20% higher pricing in GovCloud. Its identity is spec-driven development: specs, custom subagents, powers, hooks and full CLI access, with a frontier Anthropic model plus open-weight Qwen3 Coder Next and DeepSeek v3.2 available even on the free tier. The GovCloud option makes Kiro the default consideration for US public-sector teams evaluating AI coding assistants.
Open-Source AI Coding Assistants by GitHub Stars
The open-source bench is deep, and star counts give a rough popularity signal. By GitHub stars in August 2026, OpenCode leads with 172,198 (MIT-licensed), ahead of Gemini CLI at 105,104, OpenAI Codex CLI at 89,991, Cline at 62,996, Goose at 48,542 and Aider at 45,945 — the figures charted below. MightyBot’s 2026 ranking also name-checks xAI’s Grok Build, Kimi K3, Devin and Replit Agent alongside the commercial leaders.
Open source pairs naturally with API billing: you trade a subscription for direct token costs and keep full control of the toolchain — attractive for teams that already run their own platform engineering.
How to Choose Between AI Coding Assistants in 2026
There is no single best tool, but there are clear best fits. The decision comes down to workflow surface first, governance second and metered economics third. Before committing budget, it is also worth reading our companion piece on AI coding mistakes — most failed rollouts of AI coding assistants stem from usage habits, not tool choice.
AI Coding Assistants for Solo Developers
Start free: Copilot Free’s 2,000 completions and 50 chat or agent requests, Cursor Hobby, Devin Free, Antigravity’s free Individual plan or Kiro’s 50 monthly credits. When you outgrow that, the $20 shelf is where most solo developers land — Claude Pro ($17 annual) if terminal-agent work appeals, Cursor Pro if you want the richest $20 feature set, ChatGPT Plus if you already pay for it anyway. Escalate to a $100–$200 tier only once you are hitting limits weekly, because that is the point at which metered reality bites.
AI Coding Assistants for Growing Teams
Teams weigh seat pricing against admin needs. Copilot Business at $19 per user is the cheapest managed rollout and now hosts Claude Code and Codex as third-party agents, which makes it a low-risk umbrella. Cursor Teams at $40 adds SAML/OIDC SSO, privacy mode and agentic code reviews. Devin Teams’ $80-plus-$40-per-seat structure suits ten-plus seats. Standardising how AI coding assistants plug into CI/CD and review workflows matters more than the tool logo — our DevOps services team spends most of its assistant-related engagements on exactly that integration layer.
AI Coding Assistants for Regulated Enterprises
Regulated buyers should shortlist on controls, not benchmarks. Claude’s stack — managed settings that nothing can override, server-pushed policies, an OS-level sandbox on agent shell commands, SCIM, compliance API and HIPAA-ready configuration — is the most fully published. Cursor Enterprise adds pooled usage, SCIM and audit logs; Devin Enterprise offers dedicated deployment; Kiro brings GovCloud availability. Demand the same evidence from every vendor: enterprise-grade AI coding assistants should document their policy hierarchy, not just claim SOC 2.
The Two-Tool Pattern That Won 2026
The prevailing 2026 usage pattern is pairing two tools: an IDE-integrated assistant for day-to-day coding plus a terminal-based agent for heavy lifting, with heavy-usage costs typically landing at $60–$200 per month. A common pairing is Cursor Pro or Copilot Pro in the editor with Claude Code or Codex in the terminal. Budget for the pair, not the tool — and revisit the split quarterly, because as this year proved, the pricing of AI coding assistants can change faster than your renewal cycle.
Frequently Asked Questions
Are AI coding assistants worth paying for in 2026?
For most working developers, yes — but at the right tier. AI coding assistants now start at $8–$20 per month, and the free tiers (Copilot’s 2,000 completions, Antigravity’s unlimited Tab completions, Kiro’s 50 credits) are genuine trials. The $100–$200 power tiers only pay off for sustained agentic workloads, which is precisely why vendors aligned them with OpenAI’s $100–$200 estimate of real power-user consumption.
What happened to Windsurf?
Cognition acquired Windsurf on 14 July 2025 and rebranded it Devin Desktop on 2 June 2026 through an over-the-air update; plans, pricing, settings and extensions carried over unchanged, and windsurf.com now returns an HTTP 308 redirect to devin.ai. Reviews of “Windsurf” therefore describe the direct ancestor of today’s Devin Desktop.
Which AI coding assistants are cheapest to start with?
GitHub Copilot Free is the cheapest meaningful entry among AI coding assistants: $0 for 2,000 completions and 50 chat or agent requests monthly. ChatGPT Go at $8 is the lowest paid tier, and Copilot Pro at $10 with $15 in credits undercuts the otherwise universal $20 shelf. Claude Pro’s annual rate of $17 per month is the cheapest route into a frontier terminal agent.
How do AI coding assistants bill usage now?
Almost universally by metered consumption. GitHub AI Credits (since 1 June 2026) peg 1 credit to $0.01 against per-model token rates; Codex bills token consumption on ChatGPT plans (since 2 April 2026); Kiro sells credits at $0.04 as add-ons; Antigravity draws on a compute pool refreshing every 5 hours. Flat “unlimited” plans have effectively vanished from AI coding assistants at every price point.
Can Claude Code and Codex run inside GitHub Copilot?
Yes. Copilot Pro includes third-party agent integration that lets both Claude Code and Codex operate inside the Copilot experience, alongside GitHub’s cloud agents and code review. Since 14 April 2026, github.com also supports model selection for Claude and Codex agents, with a picker spanning Claude Opus 4.8 through Gemini 3.1 Pro.
Do AI coding assistants include enterprise controls at the $20 tier?
Rarely. The pattern across AI coding assistants is that identity and governance arrive one or two tiers up: Cursor’s SAML/OIDC SSO lands on Teams at $40 per user, Claude’s SCIM and compliance API on Enterprise, Devin’s dedicated deployment on custom Enterprise pricing, and Claude Code itself needs a $100–$125 Premium seat within Claude Team. Read the tier matrix before promising your security team anything.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
