Claude Code Alternatives: Codex, Open Models and Costs

Updated: 24 October 2026

There are workable Claude Code alternatives, but replacing Claude Code is three decisions, not one: the tool your developers work in, the language model (LLM) behind it, and where that model runs. Your developers can keep the same way of working with OpenAI's Codex or with a free open-source tool such as OpenCode, Cline, Qwen Code, OpenHands, Kilo Code or Goose. On Terminal-Bench 4.0, the current independent test (read 24 October 2026), the best open model solves about 4 tasks in 10 where Claude's top model solves 6.5.

Most companies want one of two outcomes. The first is to pay less, by API:

• Codex with OpenAI's GPT-6.1 Sol costs about 1/7 of Claude's cost per solved task, at about 90 % of Claude's score.
• The budget open model, GLM-5.3-Flash, costs about 1/22, at 55 % of the score.
• The top open model, GLM-5.3, costs $12.42–23.72 per solved task at the four hosts compared here, against Claude's $22.05, at two-thirds of the score. The lowest price is for a compressed build that nobody has scored.

The second is to keep the code inside the company. The model then runs on servers you control, and renting one 4-GPU server alone costs 6–10 times what ten developers spend on Claude Code. That route is chosen for control of the code, and you pay extra for it.

Either way, compare the cost per solved task, not the price per token. The line that decides the bill is cache reads: the part of each request the provider has already seen, billed at a discount. The same model at the same token price comes to $19.77 per solved task at Z.ai and $12.42 at Mistral (which serves a compressed build), and the cache price is the only line that differs.

Contents

• Why companies look for a Claude Code alternative
• How far behind Claude the open models are
• Cache reads: the line that sets the cost
• Model × where it runs, compared on one date
• The tools: Codex, open-source Claude Code alternatives, Cursor and Copilot
  • Codex vs Claude Code
  • OpenCode, the most starred open tool
  • Which models OpenCode users actually run
• The best LLM for coding, by size of hardware
• Reason 1: a cheaper Claude Code alternative by API, and a weaker one
• Reason 2: keep the code inside your network
• Checklist for the owner, and a two-week trial
• What changed in Claude Code alternatives this month
• Official sources for this comparison
• FAQ

Why companies look for a Claude Code alternative

Three concerns come up, usually together:

• The bill per developer. Anthropic's own average across enterprise deployments is $150–250 per developer a month, and heavy agent work runs past a seat's allowance. What a team actually pays is in Claude Code for business.
• Code leaving the company. Every Claude Code request sends the code it works on to Anthropic or to the cloud that hosts Claude. For most companies the commercial terms are enough; for some, code may not leave the company at all.
• One vendor for everything. The models, the prices and the usage limits belong to one vendor and change on its schedule.

The first two are the two outcomes above. The third is answered by either one: an open tool works with many models, and an open model is sold by many hosts, so you can change the host without retraining the team. The tool is the easy part to replace; the model and where it runs decide quality and cost. If the larger question is how AI should fit into your development process at all, that is the work of our vibe coding service for software teams.

How far behind Claude the open models are

Compared with Claude, the best open model, GLM-5.3 from Z.ai, scores 41.8 % on Terminal-Bench 4.0. Claude Code with Opus 5.5 scores 64.9 %. That is almost two-thirds of Claude's result: about 64 %, or 57–73 % once both margins of error are counted.

Terminal-Bench is an independent test run by Stanford, Harbor and the Laude Institute. An AI agent gets real tasks in a terminal (the text window developers type commands into) and either finishes each one or fails it. Each row is 330 attempts, and the board prints a margin of error of about ±3 points per row. Terminal-Bench 4.0, read on 24 October 2026:

"Reasoning effort" is how long the model may think before it acts. Every row except GLM-5.3-Flash ran at the highest effort the board records.

What it means. The best open model sits between Claude's July model (Opus 5) and its May model (Opus 4.8), level with Claude Fable 5 from June. In plain terms, an open model today codes roughly where Claude was in June 2026. The proprietary models gained more than 40 points in four months, and the gap has grown.

The open models were tested inside Claude Code. The board's open-model rows are Claude Code pointed at another model. We found no public board that scores OpenCode, Cline or Qwen Code with these models, so the pairing you would actually run, an open tool with an open model, is unmeasured. A provider's endpoint can also move a score: how reliably it passes back tool calls (the agent's requests to read a file or run a command), and whether the model's reasoning carries over between turns. That belongs in your trial. Using Claude Code itself with another model day to day is outside this comparison.

Two published hands-on reviews fill part of the gap:

• Builder.io, January 2026: the only published test we found with the same model in both tools (Claude Sonnet 4.5, one repository, four tasks). Quality tied, and Claude Code finished faster, 9 minutes against OpenCode's 16. It shows the two tools are comparable on one model and says nothing about open models.
• XDA Developers, April 2026: a hands-on review of OpenCode, Cline and Aider in which OpenCode "came closest" to Claude Code. It names no model for any tool and reports no measurements: a practitioner's impression.

We found no published controlled test of an open tool with an open model (searched 24 October 2026). Comparisons that call the gap closed usually rest on Terminal-Bench 2.1, where the top proprietary and open models all score about 85–89 % in their vendors' own tables, or on SWE-bench Verified, which stops in February 2026. Neither separates the models any more.

Cache reads: the line that sets the cost

A coding agent re-reads the same material on every step: your code, its instructions, the conversation so far. Providers keep that repeated part in a cache and bill it at a discount; that is a cache read. On Terminal-Bench 4.0 cache reads are about 97 % of an agent's input, so the price per token on a list tells you little.

Cost per solved task = (cache-read price × cached tokens + fresh-input price × fresh tokens + output price × output tokens, plus cache writes or storage where billed) ÷ tasks solved. The tokens spent on failed attempts are part of the cost.

A weaker model needs more attempts and more context per solved task. GLM-5.3 reads 61.2 million cached tokens per solved task; Opus 5.5 reads 36.0 million. So a cheaper cache price per token can still mean a bigger cache bill. Cache reads make up 70–80 % of an open model's bill and about half of Claude's: at Anthropic's list prices, Opus 5.5's cost per solved task is 53 % cache reads and 30 % output.

The same model costs different amounts at different hosts

GLM-5.3 is priced at $1.4 per million input tokens and $4.4 per million output tokens at both Z.ai and Mistral. Z.ai charges $0.26 per million for cache reads. Mistral charges $0.14, for a 4-bit build: a compressed copy of the model that needs about half the memory of the 8-bit ("fp8") version and has no score of its own. Per solved task that is $19.77 at Z.ai and $12.42 at Mistral, a 1.6-fold difference from the cache price alone. Amazon Bedrock lists the same model at $1.68 / $5.28 / $0.312: $23.72, about Claude's cost.

Open does not mean cheap. Qwen3.8-Max, an open-weight model sold by API at $2 / $6 per million, comes to $52.09 per solved task on Terminal-Bench 4.0, more than twice Claude's $22.05, at 27 % against 64.9 %.

The hit rate matters as much as the price

The hit rate is the share of input the provider actually serves from its cache. GLM-5.3-Flash at Z.ai costs $1.01 per solved task at Z.ai's cache price and $3.84 with no cache discount at all, 3.8 times more. Against Claude Code + Opus 5.5 at $22.05, GLM-5.3 breaks even only at a hit rate of about 95 % at Z.ai and 86 % at Mistral; at Bedrock's prices it never quite does. The board measured 98.1 % for GLM-5.3 inside Claude Code. GLM-5.3-Flash stays far below Claude's cost at any hit rate, at 55 % of its score.

Two cautions. The 97 % was measured inside Claude Code; OpenCode on an OpenAI-compatible endpoint (the standard way providers expose a model) relies on the provider caching repeated text automatically, so ask for the hit rate it sees on agent traffic. And through an aggregator such as OpenRouter, pin one host: consecutive requests sent to different hosts lose the cache.

Model × where it runs, compared on one date

Each model, each place it can run, its Terminal-Bench 4.0 score, its price including the cache-read price, the cost per solved task, and the host's own words about your data, all read on 24 October 2026. None of the 45 competing pages screened for this article that day put these side by side.

How to read the table:

1. "Per solved task" is calculated from the board's token counts, measured inside Claude Code, at each host's listed price. It is a calculation; nobody measured these at the host. For Claude, $22.05 is the board's own bill and $27.28 is Anthropic's list price applied to the board's tokens. Cache writes are not counted (Anthropic, OpenAI and Bedrock bill them). Prices are US-dollar list prices, excluding tax.
2. The score is the model's. The board's bills match the vendor's list prices; which host served the runs is not stated. A 4-bit build may lose quality and nobody has scored it. Its price edge over the 8-bit price holds only while it solves at least 87 of 330 tasks for Mistral's GLM-5.3 (the board's run solved 138), or 91 for CoreWeave's Flash (the board's run solved 118).
3. An empty score is the honest state of the evidence. DeepSeek V4.1 Flash, the most used model in OpenCode this week, has no independent row, so it gets a price and no verdict.
4. What the figure leaves out: your developers' time on failed tasks. On Terminal-Bench 4.0 Opus 5.5 failed 116 of 330, GLM-5.3-Flash 212 of 330, and someone has to finish each one. One board task is also large (about $14 per Opus attempt), so these figures compare models; they don't budget a developer's month.

Licences for the rows on your own servers: GLM-5.3-Flash is MIT and Qwen3.8-27B is Apache-2.0; GLM-5.3 has its own licence. For the API rows the provider's contract governs. The hosts are named to compare prices; naming one is no recommendation.

The tools: Codex, open-source Claude Code alternatives, Cursor and Copilot

Your developers will pick the tool they like working in, and most of these are close in features. Three things are the owner's business: what the tool sends out by default, its licence, and whether you can lock which model provider it uses.

Gemini CLI is listed for its licence and reach only; its defaults were not checked for this article.

Outdated or no longer maintained, so leave them off a shortlist: Aider (no release since August 2025), Roo Code (archived in May 2026) and Continue (read-only). Several comparisons still recommend them. Crush is source-available under the FSL licence, which is not open source.

Cursor and GitHub Copilot, proprietary tools that work with several vendors' models, are compared in Claude Code vs Cursor vs GitHub Copilot.

Codex vs Claude Code

Codex is the strongest cheaper alternative on Terminal-Bench 4.0 (read 24 October 2026), and a proprietary one: the tool is open source, the models it uses belong to OpenAI.

• Score: Codex + GPT-6.1 Sol 58.2 % against Claude Code + Opus 5.5 64.9 %, about 90 % of Claude's result.
• Cost per solved task, from the board's bills: $3.30 against $22.05, about 1/7.
• Why so much cheaper: GPT-6.1 Sol used 7.4 million input tokens per solved task; Opus 5.5 used 37.1 million. Tokens burned per solved task is the second lever on the bill, next to the cache price. The same lever explains why Claude Sonnet 5.5, priced at half of Opus per token, costs more per solved task ($35.95).
• Your data: OpenAI's API terms, quoted in the cross-check table.
• The question for your trial: whether 90 % of the score holds on your code, and what the missing 10 % costs your developers in rework.

A full Codex vs Claude Code comparison is next in this series.

OpenCode, the most starred open tool

OpenCode is MIT-licensed, with about 221,000 GitHub stars and weekly releases (v1.19.4 on 22 October 2026). Feature for feature it is the closest to Claude Code: a planning mode and a building mode, helper agents for sub-tasks, connections to outside tools, and any model through more than 75 providers or a local server.

On privacy the precise statement is: no telemetry module was found in its open-source code (read 24 October 2026). That alone does not guarantee that nothing leaves the machine. By default OpenCode also fetches its model catalogue, downloads language servers (helpers that let it understand code) and checks for updates; each can be switched off (OPENCODE_DISABLE_MODELS_FETCH, OPENCODE_DISABLE_LSP_DOWNLOAD, OPENCODE_DISABLE_AUTOUPDATE). Shared conversation pages can be disabled, and a managed configuration file can lock the list of providers for the whole company.

Which models OpenCode users actually run

OpenCode publishes its own usage data. In the seven days to 24 October 2026 the most used models were DeepSeek V4.1 Flash (112 trillion tokens), MiniMax-M3.1-Flash (39 trillion) and GLM-5.3-Flash (8.9 trillion): cheap open models. DeepSeek V4.1 Flash appeared on 10 September and took the lead within days.

It measures use, and says nothing about quality: "most used", never "best". The page does not say whether it counts every install or only traffic through OpenCode's own services, and the ranking moves weekly.

The best LLM for coding, by size of hardware

OpenCode, Cline and the other open tools run any of the models below, through a provider's OpenAI-compatible endpoint or from a server you run. On Terminal-Bench 4.0 (24 October 2026) Claude Code + Opus 5.5 leads, and the best open-weight model is GLM-5.3. If the model is to run on your own hardware, the hardware picks the tier. (H200 is Nvidia's data-centre GPU with 141 GB of memory.)

On one card, Qwen3.8-27B holds one working session of roughly 40,000–85,000 tokens of context: one developer. Several developers at once need an 80 GB-class card, by arithmetic, to be confirmed in your trial; its 4-bit build has no published score.

"Chinese alternatives" are mostly these families: GLM (Z.ai), Qwen (Alibaba), Kimi (Moonshot) and DeepSeek. GLM and Qwen hold the open rows on Terminal-Bench 4.0; Kimi and DeepSeek have no row there yet.

The licence decides who may use a model and how:

• Apache-2.0 and MIT (Qwen3.8-27B, GLM-5.3-Flash, DeepSeek V4.1 Flash): keep the notice; no other duty.
• Kimi K3 and Qwen Community: no duty for a company using the model for its own developers; extra terms only if you sell access to the model or run a very large product.
• GLM-5.3: no duty for own use; a security review by Z.ai only for model-as-a-service revenue above $10 billion a year.
• A contractor hosting the model for you: ask whose licence covers it.

Detailed pieces on Qwen and GLM are planned in this series.

Reason 1: a cheaper Claude Code alternative by API, and a weaker one

This route fits a team whose code may leave the company under a provider's contract and which wants no hardware. By API you pay per use; your own servers win on cost only at dozens of developers and with someone to run them, which is the next section.

The choice is narrower than the tables suggest (figures from Terminal-Bench 4.0, 24 October 2026, at each host's listed prices):

• Codex + GPT-6.1 Sol: about 1/7 of Claude's cost per solved task at about 90 % of its score. Today's middle ground.
• GLM-5.3-Flash: about 1/22 of the cost at 55 % of the score. A candidate for routine edits, to test in your trial.
• GLM-5.3: $12.42–23.72 per solved task at the four hosts compared here, against Claude's $22.05, at two-thirds of the score. The lowest figure is an unscored 4-bit build; at the 8-bit hosts it is about 10 % under Claude to slightly over. Its main value is as the strongest model you can run on your own servers.

Before you sign, ask the provider three questions: what its cache-read price is and what hit rate it sees on agent traffic; where requests are processed, in its contract's words; and what the contract says about training on your data and which tax applies.

Reason 2: keep the code inside your network

A closed network here means the model runs on servers you control, requests never leave your network, and the tool's own outbound calls (the tools) are switched off. It is the stronger answer for control of the code.

The setup has three parts: the tool (OpenCode, Cline, Qwen Code or OpenHands); a serving engine on your servers, the software that loads the model and answers the tool's requests; and an open-weight model, with someone who updates it and keeps it running.

RunPod's Community Cloud rents the same GPU from $3.59 an hour, but its machines belong to third-party hosts, which does not fit a closed network.

Against Claude. Ten developers on Claude Code cost about $1,500–2,500 a month at Anthropic's average. A 4×H200 server is 6–10 times that before anyone runs it, and an 8×H200 server 12–20 times. At $200 per developer, the 4×H200 rent equals the Claude Code bill of about 76 developers. These ratios compare a hardware rent with a vendor's average; they say nothing about AI making work cheaper.

What is not priced decides whether that break-even is real:

• the engineer who runs the stack, which the rent leaves out;
• how many developers one server serves at a usable speed. Memory is not the limit on 4×H200: it holds about 100 long GLM-5.3-Flash sessions. Speed is. When many sessions run at once, every step of generation reads most of the model plus each session's context, so the server's total output is capped and each developer's agent slows as more join. How it behaves with 20–50 agents at once has to be measured in your trial; we found no published test. The larger GLM-5.3 needs an 8×H200 server, and memory is then left for only about 40 long sessions, so it cannot reach a hundred-developer break-even on one server.

A closed network is chosen for the data boundary, and you pay for it.

Two checks belong in the trial. Both self-hosted models (GLM-5.3-Flash and Qwen3.8-27B) need a serving engine that reuses repeated context; without it, every turn re-processes everything the agent has read, many times the work. And your FP8 or 4-bit copy has the model's published score, not a measured one.

If this is your route, an AI platform inside your own network is the service that sizes the hardware for it and plans the pilot.

Checklist for the owner, and a two-week trial

Which route fits:

• Code may leave the company under a contract: test Codex + GPT-6.1 Sol first, and GLM-5.3-Flash for routine edits.
• Code must stay inside, and you have someone to run servers: GLM-5.3-Flash on a 4×H200 server, or GLM-5.3 on 8×H200, with the speed test below.
• Nobody in-house to run servers: stay with an API, and keep the model choice open by picking an open tool.

Before any team switches, settle these with the person who will own the setup:

• Where the code may go: outside the company under contract, or nowhere.
• The model's licence, per model.
• The tool's outbound calls and admin lock.
• The provider's terms: where requests are processed, training on your data, the contract, the tax basis, the cache-read price and the hit rate on agent traffic.
• Your sector's rules, before any external API.
• People: who owns the setup and who reviews the agent's output. The agent proposes; a person accepts.

Then run a two-week trial on your own repository. Give the same tasks to your current tool and to the candidate pair, and have your developers compare the results. Count the dull things too: how often tool calls fail on the provider's endpoint, how often the cache hits, how much rework the failed tasks leave and, for your own servers, how fast the agents stay with the whole team on. The budget model for routine edits belongs in this trial as a candidate.

What changed in Claude Code alternatives this month

The first edition of a monthly reference; each update lists what moved since the previous one.

Official sources for this comparison

All read on 24 October 2026:

• Terminal-Bench 4.0 leaderboard
• Anthropic, Claude pricing and Claude Code costs
• OpenAI API pricing and how OpenAI uses API data
• Z.ai pricing and privacy policy
• Together AI pricing and privacy policy
• Amazon Bedrock pricing, data protection and global cross-Region inference
• OpenRouter, GLM-5.3 hosts and prices (Mistral and CoreWeave rows)
• RunPod GPU pricing
• OpenCode documentation and OpenCode usage data
• Model licences on Hugging Face: GLM-5.3, GLM-5.3-Flash, Qwen3.8-27B
• Builder.io, OpenCode vs Claude Code and XDA, Claude Code against three open-source alternatives