Session 08

Claude Code·alternatives

Program
Agentic engineering
With
Arndt Voges
From
Arndt Voges Consulting
Where
Remote

The alternatives, and a look back

AlternativesLooking backClose

55 min alternatives, 25 min looking back, 5 min what comes next

0:00
Claude Code alternatives·closed and open weights, Codex, pi and OpenRouter, benchmarks
0:55
Break
1:00
Looking back·the problems log then and now, discussion
1:25
What happens next
1:30
End

Where we are

  1. 01Core concepts
  2. 02Preparing the codebase for agents
  3. 03Writing a good spec
  4. 04Feedback loops
  5. 05Workflow integration and code review
  6. 06Claude Code in depth
  7. 07Orchestration and subagents
  8. 08TodayClaude Code alternatives
Eight sessions. Today closes the program.

Seven sessions on one tool.
Today: what else exists, and how to judge it.

Today, you will go from one tool to any tool.

Specs, feedback loops and review carry over. The model and the harness are parts you can swap.

What we cover today

Closed weightsLicenseAGENTS.mdArena
CodexOpen weightspi
OpenRouterBenchmarkGPT-6Harness
SWE-bench ProTerminal-BenchZero data retention

Part one

Closed andopen weights

The elephant in the room first.

Weights

The giant pile of numbers inside the model.
Training sets them once. Every answer is computed from them.

A 27-billion-parameter model is 27 billion of these numbers. About 54 GB in one file.

The elephant in the room Anthropic and OpenAI

8 Sep 15 Sep 29 Sep ANTHROPIC OPENAI Claude Fable 5.1 1 September GPT-6 Astra 3 September Claude Opus 5.5 22 Sep · price cut GPT-6 Sol and Luna 22 Sep · price cut
The lead changes every few weeks. What you learned in this program does not. [REFRESH the week before]

Closed weights you rent the model.
Open weights: you download it.

CLOSED WEIGHTS: YOU RENT THE MODEL Your machine Claude Code, your repo The provider Anthropic, OpenAI the weights never leave context goes out, every turn Strongest models today. Your code leaves on every turn. OPEN WEIGHTS: YOU DOWNLOAD IT Your infrastructure your servers or your cloud account Claude Code or pi the harness the weights downloaded once context stays inside Your code never leaves. You run the hardware, and the quality is lower.
Open weights

The trained weights under a license.
Not the training data, not the code.

You can run and fine-tune it. You cannot rebuild it. That is why it is not open source.

The license decides

Plainly usable

DeepSeek V4 (MIT), Qwen3.8-27B (Apache-2.0), gpt-oss (Apache-2.0)

Conditions

Kimi K3, GLM-5.3, the large Qwen: revenue and branding clauses

Permission

MiniMax-M3: written permission for commercial use

Legal reads the license first.

What open weights buys you, and what it costs

Buys
  • Code never leaves your infrastructure
  • No vendor can change the terms
  • Fine-tuning on your own code
Costs
  • Hardware for a trillion parameters
  • Someone to run it
  • Quality: 53.6 against 66.0 at the top
A 27-billion-parameter model fits on one server.

Part two

Codex andthe OpenAI models

OpenAI's answer to Claude Code.

Codex

OpenAI's Claude Code.

Open-source CLI (Apache-2.0), desktop app, IDE extension, cloud tasks, GitHub reviews. Included in every ChatGPT plan.

Same concepts, other file names

Claude CodeCodex
CLAUDE.mdAGENTS.md, capped at 32 KiB
Permission modesSandbox modes plus approval policies
Skills, SKILL.mdSkills, the same standard
Subagents in .claude/agents/TOML files in .codex/agents/
claude -pcodex exec
/code-review@codex review
What you built in sessions 02 to 07 carries over, mostly by renaming.

The OpenAI models

ModelPer million tokens, in / outFor
GPT-6 Astra$10 / $50the top model
GPT-6 Sol$2 / $10coding and agents
GPT-6 Luna$0.10 / $0.50cheap and fast
All three take about a million tokens of context. [REFRESH the week before]

Before you switch: data handling.

Business and Enterprise: no training on your data by default. EU data residency depends on feature and region: the same question as in session 01.

Part three

pi andOpenRouter

Keep the harness, swap the model.

pi

A whole harness in under 1,000 tokens.

Four tools: read, write, edit, bash. No MCP, no plan mode, no permission prompts, on purpose. Reads AGENTS.md and CLAUDE.md.
OpenRouter

One API, hundreds of models.

Provider prices passed through, plus a fee on credits. Zero data retention per request; EU routing on business plans.

Swap the model, keep the harness

HARNESS MODEL Claude Code Anthropic Codex OpenAI pi open source, MIT Claude Opus, Sonnet, Fable GPT GPT-6 Astra, Sol, Luna Open weights GLM, Kimi, DeepSeek, Qwen through OpenRouter pi keeps the harness and swaps the model with one setting.
Only approved providers with zero data retention, cleared with data protection before any company code goes out.

Open-weight models worth trying

GLM-5.3

Z.ai · best open weight on the index

Kimi K3

Moonshot · strong on web front ends

DeepSeek V4

MIT license, plainly usable

Qwen3.8-27B

Apache-2.0, fits one server

MiniMax-M3

strong, restrictive license

Scores on model cards are self-reported. [REFRESH the week before]

Part four

What is out there,and what is good

Where to look, and how to read a benchmark.

How do I know what is out there?

Benchmarks: read who ran it

Anthropic's announcement

Opus 5.5 on Terminal-Bench 4.0: 66.4 %

Independent run

Artificial Analysis, same benchmark: 59.6 %

Lab-reported or independent, which harness, which version.

SWE-bench Verified is retired.

OpenAI stopped reporting it in February 2026: flawed tests and training contamination.

The benchmarks worth watching

AA Coding Agent Index

Independent, real agent harnesses, cost per task.

Terminal-Bench 4.0

Long terminal tasks, submitted by the labs.

SWE-bench Pro

Independent, several languages, lags new releases.

Arena WebDev

Human votes: front-end taste, not correctness.

Good right now harness plus model

0 20 40 60 Claude Code · Opus 5.5 66.0 $13.04 per task Claude Code · Fable 5.1 62.2 $12.39 per task Codex · GPT-6 Astra 61.6 $7.47 per task Claude Code · Opus 5 59.7 $10.79 per task OpenCode · GLM-5.3 53.6 open weights Kimi Code · Kimi K3 51.9 open weights Antigravity · Gemini 3.8 Flash 41.9 Artificial Analysis Coding Agent Index v1.5 · independent · harness plus model · as of 25 September 2026
[REFRESH the week before] Arena WebDev on the same day: Opus 5.5 first, GPT-6 Astra second.

The harness counts
as much as the model.

The index scores the pair. Compare harness and model together.

The benchmark that matters
is your own ticket.

A real ticket, the same spec and loop, three runs each. Judge the result you would merge, not the demo.

Closing

Lookingback

Eight sessions, and what comes next.

Staying current

The problems log at the start, and now.

[FILL: the pooled problems log from the start of the program against today: solved · known and handled · still open]

Where does the team stand?

What happens next

End of the program

The final report

Optional

Two sessions a month on current developments

Optional

Weekly office hours, and chat in between

Example homework, to be tailored to your team

Homework after the program

Comparison

One other harness or model

The ticket from the first homework, in Codex, or in pi with an open-weight model through an approved provider. Against Claude Code.

Ownership

One page per team

An owner for the setup, and one page: what works, what does not, what next.

Handover test

Someone new, only your setup

A person outside the team takes a ticket with only your repository setup. Note where they get stuck.

Thankyou

Arndt Voges Consulting · arndt@arndtvoges.com · arndtvoges.com