MDL.01
DeepSeek V4.1-Flash
DeepSeek V4.1-Flash for visual documents, coding and model routing: September 2026 specifications, API price comparison, open weights and migration guidance.
Open pageTechnology
AI model families.
Each on its merits.
We are not a reseller for any of these and receive nothing for recommending one. What follows is what each family is genuinely good at, what it costs, where its weights and data live, and the conditions under which we would pick something else.
- Model-agnostic
- Benchmarked per workload
New decision model
Meet Jev from TypeSafe AI.
Jev powers fast, typed decisions inside software. It adds a different role to the model catalogue: selecting tools, routing requests and assessing a bounded condition. A generative model can still produce the answer or code around it.
Explore our model-router design and the published benchmark comparison to see where that decision layer may fit.
The selection is not a ranking
There is no best model. There is a workload, a set of constraints, and a family that fits them better than the others do. Capability is rarely the binding constraint — most production tasks sit comfortably inside the reach of six or seven families, and the decision is settled by where the data is permitted to live, what latency the interface can absorb, and what a request costs at your projected volume.
That is why these pages are organised by family rather than by leaderboard position. A leaderboard tells you which model won a benchmark somebody else designed. It does not tell you that your extraction task is handled identically by a model a twentieth of the price, which is the finding that changes a budget.
Two structural questions cut across all of them, and they are worth settling before you read a single capability claim. First: open weights or hosted API? That decides whether you can run the model on hardware you control, which in turn decides whether a whole class of data-residency conversation happens at all. Second: who trained it, and who operates the endpoint? Those are different organisations for open-weight models, and the distinction matters more than most procurement processes currently reflect.
The families at a glance
Compare your options
| Criterion | Family | Weights | Runs on your hardware | Typical role |
|---|---|---|---|---|
| Anthropic Claude | Closed | No | Long-context reasoning, agentic tool use | |
| OpenAI GPT | Closed | No | Broad general capability, computer use, mature ecosystem | |
| Google Gemini | Closed | No | Very long context, native multimodal | |
| DeepSeek | Open | Yes | Frontier-adjacent reasoning at open-weight cost | |
| Llama | Open (community licence) | Yes | The default self-hosted workhorse | |
| Mistral | Open (some Apache-2.0) | Yes | EU-domiciled provider, some Apache-2.0 licences | |
| Qwen | Open | Yes | Widest size range, strong multilingual | |
| Small models (1B–14B) | Mostly open | Yes | High-volume narrow tasks at low unit cost |
Every model page
-
-
MDL.02
Claude Fable 5.1 in Production
Claude Fable 5.1 for enterprise coding and long-running agents, with September 2026 benchmark evidence, access distinctions and evaluation guidance.
Open page -
MDL.03
OpenAI GPT Integration
Where GPT models fit, what computer use and a million-token window make buildable, and the trade-offs against other frontier and open options.
Open page -
MDL.04
Gemini for Enterprise
Long-context and multimodal strengths, Vertex deployment, and where Gemini earns its place in a stack.
Open page -
MDL.05
Llama, Self-Hosted
Meta Llama on your own hardware — licence terms, ecosystem maturity and realistic serving costs.
Open page -
MDL.06
Mistral: European Open Weights
European open-weight models, Apache-licensed options and the data-residency case for using them.
Open page -
MDL.07
Qwen for Multilingual Work
Alibaba Qwen — broad model sizes, strong multilingual coverage, and the same provenance questions as any non-EU open model.
Open page -
MDL.08
Open-Weight Models
What running your own models genuinely costs, when the control is worth the capability gap, and how to decide.
Open page -
MDL.09
Small Language Models
The 1B–14B class that handles most production tasks at a fraction of frontier cost — and where it stops being enough.
Open page -
MDL.10
Embedding & Reranking Models
The models that decide retrieval quality, why they matter more than the generator, and how to evaluate them on your corpus.
Open page
How we actually decide
SELECTION PROCEDURE
-
01Constraints
residency · licence · budget
-
02Candidate set
usually 3–4 families
-
03Eval harness
your data, your task
-
04Cost & latency
at projected volume
-
05Pin a version
behind an interface
-
06 · Branch / escalationRe-run on change
deprecation is routine
Solid: selection pathDashed: the loop that runs for the life of the system
Questions
Questions about Model Families.
Do you have a preferred model?
No, and we have no commercial reason to. We hold no reseller or partner agreement with any model provider, so the recommendation is not worth anything to us beyond being right. In practice we run three or four candidates against your task and the result is frequently that two of them would work and the choice comes down to price or an infrastructure relationship you already have.
How often does the answer change?
The shortlist changes roughly every two quarters; the decision procedure does not. That is why we build systems where the model sits behind an interface — swapping a family should be an evaluation run and a config change, not a rewrite. Systems whose prompts are entangled with one model version do not survive a deprecation cycle.
Can we mix families in one system?
Usually you should. Routing straightforward requests to a small or open model and reserving a frontier model for the genuinely hard minority is the single largest cost lever available, and it typically costs nothing in quality when the routing is measured rather than guessed. See hybrid model routing for how the gate is built.
What if our data cannot leave our infrastructure?
Then the closed families are out and the decision is made among the open-weight ones, running on hardware you control. That is a real capability trade-off and we would rather state it than paper over it — though the gap has narrowed enough that for many production tasks it no longer decides anything.
Further reading
What changed this September?
Our September 2026 AI briefing connects current releases to customer decisions, including DeepSeek V4.1-Flash for visual workloads and Claude Fable 5.1 for longer coding and knowledge tasks.
Which model for
your workload?
Bring the task and the constraints. Benchmarking against your actual data beats anyone's opinion, including ours.
30-minute technical call · no deck
