Back
LLMs

Open Models: What They Cannot Do

The benchmark gap has largely closed: open models now score within a point of the closed frontier on everyday coding. What still stops you is somewhere else entirely. The licence that is not open, the indemnity nobody gives you, the long-horizon work that defeats every model on the market, and the safety layer that comes off in an afternoon.

8 min read
Updated Aug 12, 2026
QUICK ANSWER

The usual case against open models is that they are not as good.

Key Takeaways
  • This guide provides comprehensive, actionable information
  • Consider your specific workflow needs when evaluating options
  • Explore our curated LLMs tools for specific recommendations

Start by conceding the argument you were going to have

The usual case against open models is that they are not as good. On everyday coding work, that is no longer true, and pretending otherwise wastes the part of the discussion that matters.

MiniMax M2.5 scores 80.2% on SWE-bench Verified against Claude Opus 4.6's 80.8%. DeepSeek V4 Pro is at 80.6%. That is a rounding error, not a tier. GLM-5.2 ships 753 billion parameters under an MIT licence with a 1M-token context window, and the current closed leader, Opus 4.7 at 87.6%, is ahead by seven points rather than by a generation.

So the interesting question is not whether an open model can write your code. It is what happens in the parts of the job that no benchmark measures. Those are the real limits, and most of them are not about intelligence at all.

Licence
Often not open
Zero
Indemnity behind it
<10%
Tasks over four hours

1. They often cannot be used the way "open" implies

"Open weights" and "open source" are not the same claim, and the gap between them is where teams get caught. The Open Source Initiative has been explicit that Meta's Llama licence is not an open source licence, and the reasons are worth reading before you build on one.

  • A user cap. Companies above 700 million monthly active users must request a separate licence from Meta. That will never affect most readers, and it is a field-of-use restriction that disqualifies the licence regardless.
  • A no-compete clause on your own work. You may not use Llama outputs to improve any other large language model. If your plan involves distilling, generating training data, or fine-tuning something else from its output, the licence forbids it.
  • A geography clause. The multimodal Llama models are not licensed to individuals or companies based in the European Union.

None of that applies to GLM-5.2, which is genuinely MIT with no regional limits. That is the point: the licences differ enormously between models that all get called open, and the marketing does not distinguish them. Read the actual file before the architecture depends on it, because this is the one limitation on this page that a lawyer, rather than an engineer, will find for you.

2. Nobody stands behind them

Buy access to a closed model and you get a counterparty. OpenAI's Copyright Shield covers ChatGPT Enterprise and API customers against third-party IP claims arising from output. Microsoft offers the Copilot Copyright Commitment, Google indemnifies its Workspace AI features, and Anthropic commits in its commercial terms to defend customers against copyright claims from authorised use and to pay approved settlements.

Download a set of weights and there is no counterparty at all. There is nobody to indemnify you, nobody contractually obliged to defend a claim, and no negotiated agreement to point a procurement team at.

For a solo developer this is an abstraction. For anyone selling into regulated industries, enterprise procurement, or public sector, it is frequently the actual blocker, and it has nothing to do with whether the model is any good. Teams routinely discover this after the technical evaluation is finished and the model has already won.

3. They cannot finish long work, and neither can anything else

This is the limitation people expect to find, and the honest version of it is not the one they expect.

Success rates collapse as a function of task length across the whole field. Frontier models complete close to 100% of tasks that take a human under four minutes, and under 10% of tasks that take a human more than four hours. On SWE-EVO, which measures longer-horizon software evolution, the best frontier model scores 25.0%.

Read those numbers carefully, because they are not an argument against open models. They are an argument against the premise that any model currently handles long autonomous work. The open model failing your eight-hour refactor is failing at roughly the same place the closed one does.

Where a real gap does survive is the hardest end of multi-file and long-horizon agentic coding, and there the closed models remain the safer default. That is a narrower and more defensible claim than "open models are behind", and it is the one the evidence actually supports.

4. They cannot keep their own safety properties

An open model's alignment is a property of a file you now control, and control cuts both ways. Abliteration removes a model's ability to refuse by deleting a single direction from its activations, and it is neither difficult nor expensive.

If you are deploying an open model in a product, you inherit that: the safety behaviour you evaluated is not a guarantee about the weights your users end up running, and any finetune you pull from a hub may already have had it removed. Whatever refusal behaviour your use case depends on, you now own the job of verifying it, on every version you ship.

5. They cannot run themselves

The weights are the cheapest part. What you take on with them is quantisation choices, an inference engine, a serving stack, GPU capacity, and an upgrade treadmill that moves faster than most infrastructure roadmaps. We have covered both halves of that in detail: what it takes to run a model locally and what the hardware actually costs now. The short version is that a 753B model at 2-bit still needs roughly 245 GB of memory to hold, and the machine that held it comfortably has largely left the consumer market.

Renting the same model from a host removes every one of those problems and keeps the open licence, the portability, and most of the price advantage. Self-hosting is a decision about control and data, not a decision about cost.

What they can do, at what it costs

Having spent five sections on limits, the counterweight deserves real numbers. Take a 100K-input, 20K-output coding turn, which is a normal agentic exchange:

  • GLM-5.2 at $1.40 per million input and $4.40 output: about $0.23.
  • Claude Opus 5 at $5 and $25: about $1.00.
  • Claude Fable 5 at $10 and $50: about $2.00.

Roughly a quarter of the cost of a mid-frontier closed model and around a ninth of the top one, for a model scoring within a couple of points on everyday coding. That is a real advantage and it compounds hardest exactly where volume is highest: agent loops, batch jobs, evaluation harnesses, anything with a meter you would otherwise be watching.

Add the things a licence buys that a subscription cannot. The model cannot be deprecated out from under you, the weights cannot be retired on someone else's schedule, and nothing leaves your network if you choose to hold them.

Which limit is actually yours

IfYou sell into regulated industries or enterprise procurement
Indemnity is your blocker, not capability. Resolve it before the technical evaluation, not after. A closed vendor with a signed agreement may be the only answer your buyer accepts, however well the open model scores.
IfYou are building a product on top of the weights
Read the licence first. Llama restricts users above 700M MAU, forbids improving other models with its output, and excludes the EU for multimodal. GLM-5.2 is MIT with none of that. These are not interchangeable, whatever the word "open" suggests.
IfYour work is long-horizon and autonomous
Lower your expectations of everything, then prefer closed at the hardest end. Under 10% success past four hours is the industry number, not the open-source number. Break the work into shorter verifiable steps and the gap between model families mostly stops mattering.
IfYour workload is high-volume and repetitive
Open, hosted. This is where open models win outright and where the cost difference is largest. Take the licence and the price, skip the hardware.
IfRefusal behaviour is load-bearing in your product
Verify every version you ship. Alignment is removable from open weights, and a hub finetune may already have had it stripped. Evaluating the base model once is not evidence about what you are actually serving.

The summary worth keeping

Open models have closed the gap on the thing everyone measures and none of the gaps nobody measures. They match the closed frontier on everyday coding at a fraction of the price, and they still cannot give you a licence that is reliably open, a counterparty who will defend you, a guarantee about their own safety behaviour, or a solution to long-horizon work that nothing on the market has solved.

Choose on those, not on the leaderboard. The leaderboard stopped being the interesting part of this decision some time ago.

FREQUENTLY ASKED QUESTIONS
What are the real limitations of open-weight AI models in 2026?
The usual case against open models is that they are not as good.
EXPLORE TOOLS

Ready to try AI tools? Explore our curated directory:

SHARE THIS GUIDE

The usual case against open models is that they are not as good.

Share on X LinkedIn Reddit Email
Copied to clipboard