Best Open Source LLMs to Replace Claude Opus 5.5 (2026)

The best open source LLMs now trade blows with Claude Opus 5.5 on real coding work, at a fraction of the flagship price. Three weeks ago I published this league table as of September 24. Since then the proprietary tier copied the open tier’s pricing: Claude Sonnet 5.5, GPT-6.1 Sol, and Gemini 4 Argon all sit at $2/$10 per million tokens, while Claude Haiku 5.5 and GPT-6 Luna run at $0.10/$0.50. The open field answered with Step 5 Preview taking the #3 open-weights slot and Xiaomi shipping MOPD checkpoints that fix a real agentic failure mode.

I keep running these models against the paid frontier on actual coding work. The gap is narrower than the flagship price tags suggest, and the case for open source is no longer “cheaper sticker price.” It is cost per useful agent-run, downloadable weights, off-peak billing, and no rate limits.

If you still need to pick the agent before the model, start with our best AI coding tools and agents comparison and come back here for the model shortlist.

What changed since our September 2026 update

Three weeks is a long time in this market. Six things moved:

One more correction: Fable 5.1 and GPT-6 Astra at 53 used to be the highest published score on the Artificial Analysis Intelligence Index. No longer. Claude Opus 5.5 (max with fallback) sits at 58.

Why consider open source LLM alternatives?

The short version: the open field caught up on coding, and it did it while staying cheap to run.

What to evaluate

• Coding: Generate, debug, and refactor across a real repo
• Agentic work: Tool calling, long multi-step runs, not losing the thread at step 200
• Context length: How much of a codebase fits in one shot
• Cost per useful result: Not just sticker price per million tokens
• Weight availability: “Open” now ranges from MIT, to GLM’s $10B-revenue MaaS clause, to StepFun Model License 1.0 revenue terms. Read before you ship.

The proprietary bar: Claude Opus 5.5, Sonnet 5.5, GPT-6.1, Gemini 4

Here is what the open models are chasing, as of October 16, 2026.

Three operator notes on that table.

Opus 5.5 migration pain is real. Adaptive thinking is always on (depth is controlled with the effort parameter), forced tool use is gone, and thinking: disabled now returns a 400 error. If your agent stack was built on Opus 5, read the migration notes before switching.

Sonnet 5.5 is the model that competes with the open field on value. AA 56 at $2/$10 is 3 points above GPT-6 Astra while costing one fifth of its output rate. If you only want “cheap and smart” and do not care about weights, this is the proprietary answer.

The budget tier is only half a win for proprietary. Haiku 5.5 at $0.10/$0.50 with AA 43 is cheaper on input than GLM-5.3-Flash ($0.15) and scores higher (43 vs 42). But GPT-6 Luna at AA 38 sits below GLM-5.3-Flash (42), DeepSeek V4.1-Flash (39), and Kimi K3 (44) at the same sticker. So “proprietary is now cheapest” is true for Haiku and mostly false for Luna.

Also worth remembering: GPT-6-family requests over 272K input tokens reprice to $20/$75 on Astra-class billing. Long context is not free anywhere.

1. MiMo-V2.6-Pro: The new open-weights leader

MiMo-V2.6-Pro is Xiaomi’s September 22 flagship and the current top open-weights model on the Artificial Analysis Intelligence Index at 46, ahead of GLM-5.3 (45) and the Step 5 Preview / Kimi K3 pair at 44. That puts it above Grok 4.6 and Gemini 3.8 Flash, tied with Grok 4.7, and about twelve points behind Opus 5.5’s 58. It is a 1.02T sparse MoE with 42B active per token, MIT-licensed, and natively omnimodal: text, image, video, and audio in.

Two details I did not expect from a phone company’s AI lab: they published the RL training bill (about $2.62M for Pro, 30 steps in under six days and roughly 750K trajectories), and the API price stayed flat versus V2.5 at $0.435/M input and $0.87/M output.

Technical specifications

What stands out

The MOPD checkpoints: pull these, not the original release

On October 9 Xiaomi published MiMo-V2.6-Pro-MOPD and MiMo-V2.6-Flash-MOPD under MIT, plus the raw RL-stage checkpoints (MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL) and a 52-page technical report (section 5.6 covers MOPD2). The MOPD upgrade exists to fix one specific failure mode: tool-call repetition.

Xiaomi describes it plainly in their tool-call repetition diagnosis: the model “keeps issuing the same or highly similar tool calls, appearing busy while making no progress,” and the pattern “tends to go unnoticed.” That second part is the operator problem. A stuck agent that looks productive burns tokens and hours.

For self-hosters the checklist is simple: download MiMo-V2.6-Pro-MOPD (or Flash-MOPD), not the original tag. If your box is small, MiMo-V2.6-Distill-Qwen-9B (September 22, MIT) is the realistic option, and you can run local models with Ollama in Docker to keep the serving stack contained.

When to use it

• General agent work: Highest open score on AA’s agent-heavy index
• Long coding sessions: 1M context with strong DeepSWE and Terminal-Bench numbers
• Multimodal pipelines: Screenshots, video frames, audio notes in one model
• Skip if: you need a small self-host footprint. 1.02T needs a multi-GPU node even quantized; MiMo-V2.6-Flash (310B) or the 9B distill is the more realistic homelab option

2. GLM-5.3: The open coding workhorse

GLM-5.3 launched August 14 through the GLM Coding Plan, with open weights following on August 28. Same 744B-A40B base as GLM-5.2; every gain comes from post-training. The numbers show how far that goes: Terminal-Bench 3.0 jumped from 4.6 to 28.3, DeepSWE from 46.2 to 66.9, and Z.ai claims a 50% improvement on its in-house code bench. Z.ai calls it the most capable open-weights coding model, and it holds AA Intelligence 45, second among open models.

One caveat: GLM-5.3’s post-training run produced a big jump in cyber capability too. Z.ai is open about it (CyberGym 84.5, state of the art; exploitation benchmarks more than doubled versus 5.2). Useful for security research, and a reason to keep agent permissions tight either way.

Technical specifications

The Z.ai speed ladder

Since mid-September Z.ai ships four throughput tiers off the same model family. Worth knowing before you pick an endpoint:

• GLM-5.3-Flash ($0.15/$0.50): the budget tier, covered below
• GLM-5.3-FlashX (September 18, $0.37/$1.25, cache $0.09): up to ~200 tok/s
• GLM-5.3 ($1.40/$4.40): the weights-backed daily driver
• GLM-5.3-Prime (September 23, $2.80/$8.80, cache $0.56): 1.5-2x throughput, text only

FlashX and Prime are API-only. I found no weights for either on Hugging Face as of October 16. If you need downloadable weights, it is Flash or the full GLM-5.3.

Strengths

Best use cases

• Daily-driver coding agent: Strongest all-round open coding profile
• Long-horizon terminal work: TB 3.0 gains suggest it holds up where other models fall apart
• Claude Code-style stacks: Native Anthropic Messages endpoint
• Self-hosting at scale: FP8 and BF16 weights on Hugging Face and ModelScope, but 744B is cluster territory

3. DeepSeek-V4.1-Flash: The efficiency king

DeepSeek-V4.1-Flash shipped September 10 under MIT, and it is the most interesting architecture on this list. It uses a Causal Encoder-Decoder design: 552B total parameters, but only 8B active while reading your input and 16B while generating output. Combined with FP4 KV caching, its working KV cache is about 890 bytes per token, roughly a quarter of its predecessor. In plain terms: it is built to make million-token agent loops cheap to serve.

The vendor-reported numbers are aggressive: Terminal-Bench 2.1 at 90.6% and DeepSWE 74.2, above Fable 5’s 70.0 and GPT-5.6 Sol’s 73.0. Treat those as DeepSeek’s own runs until more third-party verification lands, and re-check the live leaderboard before quoting them as records. The odd part: Artificial Analysis scores it at 39 on the Intelligence Index, well below its coding results, because that index rewards breadth and V4.1-Flash is a specialist.

Technical specifications

Pricing and terms verified unchanged against the DeepSeek API docs on October 16, 2026. Peak windows are 01:00-04:00 and 06:00-10:00 UTC on weekdays.

What it does well

Platform notes worth knowing

• Current API model names are deepseek-flash and deepseek-v4-pro. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired but still accepted, billed at Flash price. Update your configs anyway.
• Anthropic-format base URL: https://api.deepseek.com/anthropic. Point Claude Code-compatible clients at this and the rest of the tooling works. Same trick applies in our Codex app with any model guide.
• Concurrency limits: 2500 for Flash vs 500 for Pro. Size agent fleets accordingly.
• Search and Fetch APIs are $0.75 per 1,000 successful requests, and DeepSeek Harness is in developer preview for agent-harness builders.

When to use it

• Always-on agents on a budget: $0.30/$1.20 peak, half that off-peak
• Terminal and repo-level coding: Where its benchmark profile is strongest
• Self-host experiments: MIT weights plus the most cache-efficient serving story here
• Skip if: you need broad general-knowledge work; its AA Index shows the trade-offs of specialization. A V4.1-Pro is coming for that.

4. GLM-5.3-Flash: The value monster

GLM-5.3-Flash released August 26 under MIT, and it quietly rearranged the budget tier. 320B total, 18B active, with a hybrid attention stack that interleaves 34 KDA linear-attention layers with 11 sparse-attention layers in a 3:1 pattern. That cuts attention compute roughly 3x and KV cache roughly 4.4x versus GLM-5.3. It is also the first natively multimodal GLM-5 model, with image and video input.

The punchline is the price: $0.15/M input and $0.50/M output, with $0.03 cached. It scores 84.3% on Terminal-Bench 2.1, 63.4 on DeepSWE, and holds AA Intelligence 42. Fun detail: before launch it circulated on OpenRouter and OpenCode as the anonymous “Ox Alpha” model people were trying to identify.

Technical specifications

What it does well

Best use cases

• Default model for everything cheap: Still my budget pick for open weights; M3 has its place via the MiniMax Token Plan if you want a flat subscription instead of per-token billing
• Self-hosted coding stack: MIT license, vLLM/SGLang support, one node instead of a cluster
• Multimodal on a budget: Screenshot-to-code and video understanding without K3 pricing
• High-frequency agent loops: Cache hits at $0.03/M make warm sessions nearly free
• Need more speed? GLM-5.3-FlashX runs up to ~200 tok/s at $0.37/$1.25 (API-only)

5. Kimi K3: Still the largest open weights

Kimi K3 is no longer in the top three on the AA index. AA Intelligence 44 now puts it tied with Step 5 Preview, effectively fourth on AA’s ordering: MiMo-V2.6-Pro 46, GLM-5.3 45, Step 5 Preview 44, K3 44. What has not changed: at 2.8T parameters it remains the biggest open release, and its coding numbers still hold up. Terminal-Bench 2.1 at 88.3%, DeepSWE 67.5, SWE-Marathon 42.0, FrontierSWE 81.2, and BrowseComp 91.2 for web research agents.

Since September 18 it is also generally available on Amazon Bedrock, which matters if your org wants the weights-model story inside an AWS security boundary with prompt caching. That is a September claim I have not been able to re-verify this round; check the Bedrock docs if it is load-bearing for you.

The pricing caveat from the original version of this guide still stands: $3/M input and $15/M output is frontier-adjacent money. Cache hits at $0.30/M are what keep it sane in warm coding loops. If your cache hit rate is bad, MiMo-V2.6-Pro does comparable work for a seventh of the output price. And if you are routing through OpenRouter instead of going direct, moonshotai/kimi-k3 bills about $0.304 input / $12.53 output / $0.21 cache as of October 16, roughly 17% less on output than the direct rate.

Technical specifications

What stands out

When to use it

• Hard coding and agent marathons: FrontierSWE and SWE-Marathon leadership among open models
• Web-heavy research agents: BrowseComp 91.2
• AWS shops: Bedrock with caching since September 18
• Skip if you are price-sensitive: $15/M output is the highest here by a wide margin

6. Qwen3.8-Max and the open 2.4T weights

Alibaba is running Qwen 3.8 as two products wearing one name. Qwen3.8-Max is the hosted API at $2/$6 per million tokens, live since August 3 and refreshed September 2 as Qwen3.8-Max-0902. That snapshot debuted at #1 on Arena’s Code Arena WebDev leaderboard with 1,691 points, edging Claude Opus 5 by 3 points and Kimi K3 by 17. Its Terminal-Bench 3.0 and ProgramBench scores more than doubled versus the launch build. If your work is frontend and “vibe coding,” this is the strongest open-ecosystem option right now.

The downloadable half is Qwen3.8-2.4T-A95B, released August 12: 2.4T parameters, 95B active, hybrid linear-plus-full attention, 262K native context extensible to 1M, under a custom Qwen license. Community MXFP4/NVFP4 quants squeeze it to about 1.2TB, which fits a single 8-GPU B300 node. One catch: the September 0902 post-training is API-only so far, so the weights lag the hosted model. Confirmed October 16: the Qwen org shows no new 3.8 weights since August (only Qwen3Guard-Stream safety models and Qwen-Image-2.1).

Technical specifications

That AA split matters. The API model you can call today scores 45, which edges GLM-5.3. The weights you can download still score about 40. If your eval says “Qwen3.8-Max,” make sure you know which one it ran.

What it offers

When to use it

• Frontend and design-to-code: The WebDev arena lead is the whole argument
• API convenience plus a weights exit: Start on QwenCloud, keep the option to self-host the 2.4T
• Self-host frontier-class: If you have the node for it, this is the second-largest open model after K3
• Skip if: you need the September improvements on-prem today; the 0902 checkpoint is hosted-only for now

For companion coverage of the smaller Qwen models and how they behave inside coding agents, see our Qwen models for AI coding agents guide.

7. Step 5 Preview: The new #3 open model

StepFun released Step 5 Preview on September 18 and it went largely unnoticed until the AA index caught up. It is a 600B total / 27B active MoE with 1M context, and AA Intelligence 44 puts it at #3 among open-weights models on AA’s own ordering, a hair ahead of Kimi K3. On active-parameter efficiency that is the best deal in the open top five: 27B active is a third of MiMo’s 42B.

Weights are on Hugging Face as stepfun-ai/Step-5-Preview, with BF16 and NVFP4 official variants plus community GGUF and MLX quants. Downloads are still low (hundreds to low thousands), so treat this as early.

Technical specifications

Open source LLMs vs Claude Opus 5.5 and GPT-6.1: side-by-side comparison

Numbers as of October 16, 2026. “n/a” means not published or not comparable on that exact suite. DeepSWE and Terminal-Bench columns mix vendor and third-party runs; single-source numbers get less trust. Read the table like this: at the $2/$10 tier the argument is capability per dollar, and at the $0.10/$0.50 tier it is weights and self-hosting.

*Opus 5 figure where Opus 5.5 has no published DeepSWE run; Anthropic rates 5.5 at roughly Fable 5.1 level. †AA scores the 0902 API snapshot at 45 and the open 2.4T-A95B weights at about 40. Kimi K3 prices shown are direct list; OpenRouter bills $0.304/$12.53. MiniMax M3 is the OpenRouter listing, see the pricing footnotes below. DeepSeek input is $0.30 peak / $0.15 off-peak.

Chart comparing open source LLMs against Claude Opus 5.5, Sonnet 5.5, and GPT-6.1 on benchmark scores and cost per million tokens, October 2026

How to run these open source LLMs (OpenRouter, direct API, or self-hosted)

Step 1: Choose your access path

Step 2: Environment (OpenRouter example)

# Install OpenAI SDK
pip install openai

# Set environment variables
export OPENROUTER_API_KEY="your_api_key_here"
export OPENROUTER_BASE_URL="https://openrouter.ai/api/v1"

Verify the key before you write any code:

curl -s https://openrouter.ai/api/v1/models \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" | head

You want JSON back. If you get an HTML error page or 401, fix the key or the network path first.

Step 3: First call, two models side by side

import openai

client = openai.OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="your_openrouter_api_key"
)

def ask(model: str, prompt: str) -> str:
    response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}]
    )
    usage = response.usage
    print(f"{model}: {usage.prompt_tokens} in / {usage.completion_tokens} out")
    return response.choices[0].message.content

prompt = "Create a Python web scraper for product prices"

# The open coding workhorse vs the new proprietary budget tier
print(ask("z-ai/glm-5.3", prompt))
print(ask("anthropic/claude-haiku-5.5", prompt))

Same call in curl if you just want to see the shape of the response:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"z-ai/glm-5.3","messages":[{"role":"user","content":"Create a Python web scraper for product prices"}]}'

Note on DeepSeek: the current API model name is deepseek-flash (serving V4.1-Flash), and if you want Anthropic Messages format the base URL is https://api.deepseek.com/anthropic. Legacy names like deepseek-v4-flash still resolve, billed at Flash price, but do not build new configs on them.

Step 4: Match context to the job

Step 5: Verify it works

Three checks that take five minutes and save an afternoon:

1. The response has content and response.model matches what you requested. Providers silently substitute models more often than they admit.
2. Log the usage block and price one real task. Multiply prompt_tokens and completion_tokens by the list rates. Compare that against the same task on Haiku 5.5 or Opus 5.5. Your cost per useful result is the number that matters.
3. Run the same prompt through two models and diff the output quality. This is “measure in your domain” made executable. Do it on a task you know the answer to, not a toy prompt.

Self-hosting notes

If you are going to serve weights yourself, the practical picks as of October 2026:

Two ops notes.

Rollback is boring, which is the point. A self-hosted model is weights plus an API process. Pin a checkpoint tag in your deploy config, and reverting a bad upgrade is redeploying the previous tag. There is no migration, no data format change.

Cost math before you buy iron. A 320B-class model (GLM-5.3-Flash) needs one multi-GPU node. A 2.4T-class model (Qwen3.8, K3) needs an 8-GPU node class. At the volumes in the next section (10M in + 10M out per month) API spend is $6.50 for GLM-5.3-Flash and $80-$180 for Qwen3.8-Max and Kimi K3. Hardware wins only when you need data locality, sustained much-higher volume, or cache-heavy workloads you would rather not meter. For the small-footprint end of the scale, see best mini PCs for local LLMs.

Cost breakdown: what these cheap AI coding models really cost

Illustrative monthly cost at 10M input + 10M output tokens, API list rates as of October 16, 2026. Real bills move with cache hits, plans, batch discounts, and off-peak timing. Opus 5.5 is the baseline.

Monthly cost comparison (10M in + 10M out)

The punchline changed since September. At $6/month-class spend the proprietary tier now matches open sticker price: Haiku 5.5 and Luna both land at $6 on this math. You pick open models anyway for weights, self-hosting, off-peak and batch pricing, and no rate limits. And the top of the open field (MiMo 46, GLM-5.3 45) still costs $13-$58 at this volume while beating Haiku by 2-3 AA points and Luna by 7-8. The trade: $6 buys you 43 (Haiku) or 38 (Luna) with no weights, or $6.50 buys you 42 with MIT weights and cache hits at $0.03.

Two more footnotes on that table. GPT-6 Astra and Fable 5.1 at $600 buy the top of the index only if your workload needs AA 53-58. Most coding-agent work does not. And Kimi K3’s $15/M output means its “cheap” reputation only holds when your input cache hit rate stays high.

If you would rather pay a flat monthly rate and let someone else do the routing, the OpenCode Go vs ClinePass for cheap access to DeepSeek, Qwen, and Kimi breakdown covers the subscription side of the same models.

Tips and common mistakes

What works

What to avoid

• Routing every task to the most expensive model by default. At $20/M output, Opus 5.5 on bulk refactors lights money on fire
• Trusting vendor-reported benchmarks without independent runs (DeepSeek V4.1-Flash’s Terminal-Bench numbers are self-reported so far)
• Letting agents run without permission scoping, especially given how fast cyber capability is climbing in these releases. Run AI CLI tools safely in Docker or Podman instead of on your home directory
• Ignoring the license fine print: MIT is clean, GLM-5.3 has a $10B MaaS clause, StepFun Model License 1.0 has revenue terms, Qwen, Kimi, and Muse use custom licenses
• Missing tool-call repetition loops. If an agent looks busy but the file tree never changes, you are burning tokens on repeated tool calls. Swap to an MOPD checkpoint or add loop detection in the harness. Xiaomi’s diagnosis post describes the pattern well

What’s coming next

• DeepSeek-V4.1-Pro: The family flagship. Still absent from Hugging Face and the API as of October 16; deepseek-v4-pro still serves DeepSeek-V4-Pro-0813 at $0.66-$1.32 in / $1.98-$3.96 out
• Refreshed Qwen3.8 weights: The 0902 post-training is API-only today. Verified October 16: no new 3.8 weights on the Qwen org since August. An updated checkpoint would matter for self-hosters
• Tencent Hy4: 770B Apache-2.0 preview, still tencent/hy4-preview on OpenRouter at $0.751/$2.251 and on Hugging Face with FP8 and community GGUF quants. No GA
• Step 5 GA and pricing: If StepFun prices the final model under K3’s $3/$15, it becomes the default hard-task open model
• Meta’s Muse line: Muse Wonder (AA 52, October 12) is the proprietary flagship; Muse Spark Open (October 15, 816B/42B, AA 39, Muse Model License) is the open release with day-one NVFP4 and MLX quants. Read the Muse license revenue terms before recommending either commercially
• Longer shots: GLM-5.3-Prime and FlashX weights (both API-only today), Inkling from Thinking Machines (975B/41B, Apache-2.0, AA 25), and Reflection’s Beam 1.0 (501B/23B, Apache-2.0, AA 30). Interesting licenses, not coding contenders yet

Which open source LLM should you pick?

MiMo-V2.6-Pro if you want:

• The highest open-weights score (AA 46) with MIT weights
• Strong agentic numbers (AutomationBench 53.1, DeepSWE 71.9)
• Text + image + video + audio in one model
• $0.44/$0.87 pricing, ~134 tok/s
• Pull the MOPD checkpoint for agent work

GLM-5.3 if you need:

• The strongest open coding profile (TB 3.0 28.3, FrontierSWE 78.1)
• A proven daily driver with Coding Plan subscription options
• Anthropic-compatible endpoint for Claude Code-style tools
• Predictable $1.40/$4.40 pricing

DeepSeek V4.1 Flash if you want:

• Vendor-reported 90.6% Terminal-Bench 2.1 and 74.2 DeepSWE (verify against the live leaderboard)
• MIT weights and the cheapest serious agent serving
• Off-peak pricing that halves an already low bill, weekends included
• 384K output budget for long generations

GLM-5.3-Flash if you want:

• $0.15/$0.50 list pricing with MIT weights
• Native image + video input
• A 320B model that fits one node
• 84.3% TB 2.1, four AA points behind its flagship sibling

Kimi K3 if you want:

• The largest open weights (2.8T)
• BrowseComp 91.2 research agents
• AWS Bedrock deployment with prompt caching
• You can live with $15/M output and lean on cache hits

Qwen3.8-Max if you care about:

• #1 on Code Arena WebDev for frontend work
• $2/$6 API pricing with 1M context
• A 2.4T open-weights path when you are ready to self-host

Step 5 Preview if you want:

• The new #3 open score (AA 44) with only 27B active parameters
• 600B weights with BF16/NVFP4 and community GGUF quants already out
• You can live with a preview checkpoint and a custom license while pricing is TBD

My ranking for coding-agent work right now: GLM-5.3 as the default, MiMo-V2.6-Pro (MOPD) when the task is long or multimodal, GLM-5.3-Flash for volume when you need weights and Haiku 5.5 when you do not, DeepSeek V4.1-Flash for terminal-heavy loops, and Kimi K3 or Qwen3.8-Max when their specific strengths (research, frontend) match the job. Step 5 Preview joins the hard-task list the day it gets real pricing.

For agentic harnesses specifically, the best open source models for OpenClaw shortlist filters this same set by tool-calling reliability.

One sentence on when to skip open entirely: Claude Sonnet 5.5 at AA 56 and $2/$10 is the value threat, and if you do not need weights, off-peak billing, or a self-host exit, it is a defensible default. I still pick open for anything long-running, because pinning a checkpoint and serving it yourself is the only version of “this model will still exist next quarter” you control.

FAQ