The 9 Best Claude Code Alternatives in 2026 (Tested and Priced)

Why you are probably looking

You are not leaving because the output is bad. You are leaving because of the meter.

Claude Code is included on every Claude plan, right down to the free one, and that sounds generous until you read the next line on Anthropic own pricing page: your terminal work and your chats draw from a single pool. Spend Tuesday afternoon on a big refactor and you will find your writing rationed on Thursday.

Anthropic does publish the numbers, which is more than most. On Pro you get roughly 10 to 40 prompts every five hours. On Max 5x, 50 to 200. On Max 20x, 200 to 800. One catch worth knowing before you quote those figures at anyone: the support article carrying them still describes model generations that have since been replaced, so treat them as the right shape rather than the current truth.

What it actually costs you

Here is the number nobody else on this topic quotes, and it comes from Anthropic itself. In its own cost documentation, Anthropic says the average is around $13 per developer per active day, and $150 to $250 per developer per month, with 90% of users staying under $30 a day.

Read that against a $20 Pro plan and the picture clarifies. If you are a light user, the subscription is a bargain. If you are running agents most of the day, you are consuming several times what you pay, which is exactly why the limits exist and exactly why they bite.

Three things every other guide on this page gets wrong

Before you take advice from anywhere, including here, check the dates. Three facts changed this year and most published guides have not caught up.

• Gemini CLI is no longer free for you. Google announced on 19 May 2026 that it would stop serving requests for Google AI Pro, Ultra and free users on 18 June 2026. Only Standard and Enterprise licences kept it. Antigravity CLI is the replacement. Guides still calling Gemini CLI the most generous free tier are recommending a door that is closed.
• Claude Code does not read AGENTS.md. Anthropic says so plainly: it reads CLAUDE.md, not AGENTS.md. If you have seen a guide claim it falls back, that guide is wrong and the mistake is spreading.
• OpenCode can no longer use your Claude subscription out of the box. Following a legal request from Anthropic, OpenCode removed its Anthropic auth plugin in March 2026. Its docs now state that using a Pro or Max subscription there is not supported. Plan around that before you build a workflow on it.

What the benchmarks actually say

If you only take one thing from this section: the harness moves the score more than the model does. That is not an opinion, it falls straight out of the leaderboard.

These are the official Terminal-Bench 2.1 results, 89 tasks, with the error bars the leaderboard publishes and almost nobody quotes.

Agent	Model	Score	Read it as
Claude Code	Fable 5	83.8% ±1.2	Joint first
Codex	GPT-5.5	83.1% ±1.1	Joint first
Terminus 2	Fable 5	80.4% ±1.2	Same model, worse harness
Cursor CLI	Grok 4.5	79.3% ±1.5	Close behind
Claude Code	Opus 4.8	78.9% ±1.3	Three-way tie
Codex	GPT-5.6 Terra	78.4% ±1.3	Three-way tie
Claude Code	Sonnet 5	74.6% ±1.6	The cheap model, still strong
Claude Code	GLM-5.1	58.7% ±1.2	What routing out costs you

Three things fall out of that table that should change how you shop.

Codex and Claude Code are tied. A 0.7 point gap with error bars of ±1.2 and ±1.1 is not a gap. Anyone telling you Claude Code is definitively number one is over-reading 89 tasks.

The wrapper matters more than the brain. Run the same Fable 5 model in Claude Code and you get 83.8%. Run it in Terminus 2 and you get 80.4%. That 3.4 point harness difference is bigger than the difference between the top two models. Whichever tool you pick, the tool itself is doing real work.

Cheap models cost more than they save, sometimes. Claude Code with GLM-5.1 drops to 58.7%. If you are considering pointing Claude Code at a budget model to dodge limits, that is the trade you are making — 25 points of capability.

One more note if you compare leaderboards. Artificial Analysis and Vals both re-ran Terminal-Bench 2.1 in their own sandboxes and got numbers around four points apart from each other on identical models. The ordering held; the absolute figures did not. Treat any single benchmark number as a range.

Claude Code alternatives compared at a glance

Prices and limits read from each vendor own page in August 2026. Where a vendor refuses to publish a number, we say so rather than filling the gap with a guess.

Tool	Licence	Model choice	Entry price	Free tier
Codex CLI	Apache-2.0	OpenAI on plan auth	$8 per mo (Go)	Limited
OpenCode	MIT	60+ models	Free + provider costs	Yes
Cline	Apache-2.0	Anything, incl. local	Free + provider costs	Yes
Cursor	Proprietary	Frontier + in-house	$20 per mo	Hobby tier
GitHub Copilot CLI	Proprietary	Model picker	$10 per mo	2,000 completions
Antigravity CLI	Proprietary	Gemini + Claude 4.6	$0 individual tier	Yes, limits unpublished
Charm Crush	FSL-1.1-MIT	Anything, incl. local	Free + provider costs	Yes
Goose	Apache-2.0	15+ providers	Free + provider costs	Yes
Amp	Proprietary	Routed for you	$20 per mo	Undocumented
Claude Code	Proprietary	Anthropic only	$17 per mo (annual)	Yes, on the free plan

One licence note, because it trips people up. Crush is FSL-1.1-MIT, which is source-available rather than open source — you can read and modify it, but it is not OSI-approved and it converts to MIT only after two years. If your company has a policy on this, that distinction is the one your legal team will ask about.

The ones worth your time, reviewed

1. Codex CLI — start here if you want like-for-like

This is the swap that changes least about how you work. Same terminal-first shape, statistically the same benchmark performance, and two things Claude Code will not give you.

First, price: ChatGPT Go is $8 a month against Claude Pro at $20. Second, and more useful day to day, OpenAI publishes exact limits per model per five-hour window — 10 to 100 messages on Sol, 25 to 200 on Terra, 250 to 2,000 on Luna for a Plus account. You can plan a working day around numbers like that. You cannot plan around a limit you discover by hitting it.

What you give up: hooks drop from 17 lifecycle events to 3, output styles do not exist, and your per-command permission rules have no equivalent. More on that in the migration section, because it is the part that will actually cost you an afternoon.

2. OpenCode — the efficiency argument

The headline finding here is worth your attention. An independent measurement using a logging proxy between each harness and the model endpoint found Claude Code sending roughly 32,800 tokens on a first turn where OpenCode sent about 6,900 — a 4.7x difference before you have asked for anything real.

Be careful how you read that. Over a full write-run-test-fix loop the two came out near parity, because OpenCode used more requests to get there. So the saving is real at the start of a session and shrinks as work gets longer. If you open a lot of short sessions, it matters a great deal.

MIT licensed, 60-plus models, and the most genuinely portable of the majors. Just go in knowing the subscription route is closed and you will be paying providers directly.

3. Cline — the one that shows its work

If what unsettles you about agentic coding is that it acts and then tells you, Cline is the antidote. Plan mode produces the whole plan first; Act mode executes it with per-step approval. You watch every file it intends to touch.

Apache-2.0, no seat fee, no platform cut — you pay only for inference. It talks to Anthropic, OpenAI, Google, OpenRouter’s couple of hundred models, Bedrock, Vertex, Groq, Ollama and LM Studio. For a team that needs to justify what an agent did, the audit trail alone earns its place.

4. Cursor — when you want to see the diff

Different shape of tool, and worth considering if your real complaint is legibility rather than limits. Terminal agents tell you what they did; Cursor shows you a multi-file diff you accept or reject hunk by hunk.

Hobby is free with limited agent requests, Pro is $20 a month, Teams Standard is $40 per user. Cursor does not publish dollar figures for Pro+ or Ultra on its pricing page — it publishes multiples of Pro instead, so check inside your account before you assume what an upgrade costs.

5. GitHub Copilot CLI — cheapest way in, and honest about limits

Ten dollars a month for Pro, and the free tier is the only one in this guide with a published number attached: 2,000 completions and 50 premium requests a month. That is a small allowance, but it is a knowable one.

Watch the multipliers rather than the sticker price. Every premium request is not equal: Haiku 4.5 costs 0.33 of a request, Sonnet costs 1, Opus costs 3, and GPT-5.5 costs 7.5. Letting Copilot auto-select models earns you a 10% discount on the multiplier. If you burn through quota faster than you expected, the model picker is almost always why.

6. Antigravity CLI — the surprising Claude route

Google moved Gemini CLI users here in June. The reason to care has nothing to do with Gemini: Antigravity individual tier is $0 and includes access to Claude Sonnet and Opus 4.6 alongside Google models. You can run Claude models in an agent without an Anthropic subscription.

The catch is that Google will not tell you the limits. Its own pages say only basic weekly rate limits on the free tier and more generous limits on paid, with quotas refreshing every five hours or weekly. No figures. After the Gemini CLI sunset, that vagueness is worth factoring into how much you build on it.

What breaks when you switch

This is the part nobody writes and the part that will eat your afternoon, so here it is concretely. Moving from Claude Code to Codex, roughly by how much pain each layer causes.

What you have	Where it goes	How much work
CLAUDE.md	AGENTS.md	Rename, done
MCP servers (.mcp.json)	TOML config block	Mechanical reformat
Skills folder	.agents/skills/	Directory rename
Subagents	TOML agents	Reshape, different semantics
Slash commands	Prompts and skills	Re-author by hand
Hooks	3 lifecycle events	Rethink from 17 down to 3
Permission allowlists	Sandbox + approval policy	Breaks — rethink the intent
Output styles	Nothing	No equivalent

Two practical notes that will save you time.

Do not replace your instruction file, bridge it. AGENTS.md is the closest thing this space has to a standard — it is stewarded under the Linux Foundation and supported by 30-plus tools. Claude Code is not one of them, but you can keep one source of truth by writing your content in AGENTS.md and putting a single @AGENTS.md import at the top of CLAUDE.md. Every tool then reads the same file and you stop maintaining two.

Migration runs both ways. Recent Claude Code versions ship an /import command that pulls configuration from Codex and Gemini CLI — instruction files, MCP servers, commands, subagents and skills — with a dry-run flag so you can see what it would do first. Given what happened to Gemini CLI in June, that is a genuinely useful escape hatch in the other direction.

How to cut the bill without changing tools

Before you switch anything, try this list. Most of it comes from Anthropic own cost documentation, and in our experience the first two items alone solve the problem for a lot of people.

• Clear between tasks. Stale context is billed on every subsequent message. Clearing costs nothing. Compacting, by contrast, is itself an expensive request because it has to read the conversation it summarises.
• Stop using your best model for everything. Sonnet 5 is $2 and $10 per million tokens; Fable 5 is $10 and $50. That is five times the price for work that is usually routine. Reserve the expensive model for architecture and hard debugging.
• Watch your cache lifetime. Prompt caching lasts an hour on a subscription but drops to five minutes once you start drawing on usage credits, and cache reads bill at a tenth of input price. Claude Code flags cache misses in its usage view when they pass 10% of recent usage — if you see that, it is telling you something expensive.
• Keep your instruction file short. It loads every single session. Under 200 lines is a good target; move anything situational into skills, which load only when needed.
• Filter noisy output before it reaches the model. A hook that greps test output for failures rather than passing the whole log through can turn tens of thousands of tokens into hundreds.
• Turn thinking down. Thinking tokens bill as output and default budgets run high. Lower them and most everyday work is unaffected.
• Be careful with agent teams. They use roughly seven times the tokens of a standard session. Brilliant occasionally, ruinous as a default.

If you are sizing this for a team rather than yourself, Anthropic publishes per-user rate guidance that is genuinely rare: roughly 200k to 300k tokens per minute per user for a team of five, falling to 10k to 15k for teams over 500. Plan capacity from those, not from your own single-seat experience.

How to choose

• You want the same thing, cheaper — Codex CLI on ChatGPT Go, $8 a month, and you get published limits into the bargain.
• You want to stop renting — OpenCode or Cline with your own API key. You pay providers directly and nobody can change your entitlement.
• You need to review before anything runs — Cline plan mode, or Cursor if you would rather read a diff than a log.
• You have no budget at all — Copilot Free for a known small quota, Antigravity for an unknown larger one, or a free harness pointed at a cheap model.
• You want Claude models without an Anthropic plan — Antigravity CLI, oddly enough.
• You run local models — Crush or Cline, both of which talk to Ollama and LM Studio without ceremony.
• Your team needs repeatable workflows — Goose, where recipes are files you can commit and share.
• You do not want to think about routing — Amp, which decides for you and does not mark up the provider price.

Is switching worth it?

Stay where you are if your setup is deep — a tuned CLAUDE.md, custom hooks, permission rules you trust — and your real problem is a busy fortnight rather than a permanent ceiling. The migration table above is the cost of leaving, and for a heavily customised setup it is a day, not an hour.

Switch if you are on Pro and hitting limits weekly, if you need a model Anthropic does not offer, if your organisation needs an audit trail Claude Code does not produce, or if you have worked out that at $150 to $250 a month of consumption you would rather pay providers directly and keep the control.

Or do what most experienced people quietly do: run two. Keep Claude Code for the hard problems, put a cheap harness on a cheap model in front of the routine edits, and stop paying frontier rates for work that never needed them. That single change usually saves more than switching vendor ever would.

If it is the chat assistant rather than the terminal you are reconsidering, our Claude alternatives guide covers that side, and the developer tool alternatives hub has the rest of the category.