Model selection
Reproducible comparisons across providers and open models, on your use cases, not on a paid benchmark.
Slash AI Lab
The AI lab behind every decision
We evaluate models and architectures with client data (in Spanish, French and English) before committing cost or trust. What fails evaluation doesn't ship.
Models OpenAI · Anthropic · Gemini · DeepSeek · GLM · Qwen · Mistral · open sourceAxes coding · reasoning · language · agents · costRule Evaluation, not hype
In short
Slash AI Lab is the AI evaluation lab of Slash Digital Lab, in Bogotá: we test models (OpenAI, Anthropic, Gemini, DeepSeek, GLM, Qwen, Mistral and local open source) on the client's real prompts before committing cost or private data.
Lab rule: no model is chosen by hype. It's chosen by evaluation on the client's prompts, cost, data residency and observed failures.
What we deliver
What the lab evaluates
Evaluation axes: business Spanish, French, code, agents, cost, latency and privacy.
Language quality
Your operation's Spanish and French are evaluated as hard as English; nuances matter.
Agents & tooling
MCP, function calling, memory and human oversight: we measure what breaks when the agent errs.
Cost, latency, privacy
Inference budget, p95 response time and data residency: cloud, VPC or local inference.
How we work
How an evaluation runs
Collection
We gather real or representative prompts and success criteria with your team.
Runs
Candidates run in cloud or local; we measure quality, cost and latency.
Report
Results per axis with examples: a clear, reproducible recommendation.
Watch
The market moves every quarter; when it does, we re-run.
30fixed prompts per evaluation
3languages evaluated: ES · EN · FR
0hype: only reproducible decisions
5evaluation axes
Scope and deliverables
The method, in writing
| Evaluation | 30 fixed prompts, 5 engines, 3 languages (methodology published in /insights/). |
|---|---|
| Data residency | Standard cloud, VPC or local inference depending on compliance. |
| Report | Comparison per axis: language, code, agents, cost, latency, failures. |
| Reproducibility | Prompts, versions and criteria documented to repeat the run. |
| Cadence | Quarterly re-evaluation or on relevant market moves. |
| Publication | Findings open in /insights/; client prompts never. |
Method
The lab's rules
The same in every publication and every project.
Methodology before numbers
We publish the method before the results; no finished run, no number.
Real prompts
We evaluate on the client's tasks, not toys: 30 fixed, versioned prompts per case.
Three languages
Business Spanish, English and French: that's where models break.
Declared limitations
Every analysis says what it doesn't measure, with its date, versions and linked sources.
Publications
Published criteria, not opinions
Models
Claude Opus 5.5: what changes for businesses
Launched September 22, 2026: cheaper than Opus 5, Fable 5.1-level results per Anthropic. Pricing, benchmarks, breaking changes and when to use it.
ModelsHow we choose a model
Language, code, agents, cost and privacy: the evaluation we run before every deployment, with 30 real prompts per client.
ModelsThe AI model map, October 2026
Claude Opus 5.5 and Sonnet 5.5, GPT-6 Astra, Gemini 4 Argon, Grok 4.7, Muse Spark: who leads in October 2026, what each price tier costs and what to test.
Agents and engineeringEvals: how to test an LLM system
Why eyeballing fails: a golden set in Spanish and French, calibrated LLM judges, CI regression gates, tooling and what running evals costs.
Local AI and open weightsOpen-weight models in 2026: which to use, under which license
Kimi K3, Qwen3.8, DeepSeek V4.1, GLM-5.3, Gemma 4, gpt-oss, Mistral: sizes, licenses, gap to the frontier and how to host them without legal surprises.
Agents and engineeringFine-tuning vs RAG: the criterion
When RAG, when fine-tuning and when a hybrid: the decision order before paying for any training.
Frequently asked questions
The questions we hear often
Which models do you evaluate?
The ones that can earn a slot: OpenAI, Anthropic, Gemini, DeepSeek, GLM, Qwen, Mistral and local open source. The list moves with the market.
Do you publish results?
Yes: method and findings go to /insights/. Client prompts are never published.
Is going through the lab mandatory?
Recommended, not mandatory. For sensitive-data or minority-language projects, evaluation avoids months of drift.
Do you offer local inference?
Yes: open models in your VPC or own hardware when compliance requires it.
Can we evaluate a single use case?
Yes: one focused run with your prompts and a reproducible report.
Do you accept sponsorships?
No. Slash research isn't for sale.
Related capabilities
Where to next
AI
Artificial intelligence
Applied AI for real operations: agents, RAG, vision, voice and automation.
Learn more→ Fine-tuningFine-tuning & RAG
When to tune the model and when to give it knowledge. Judgement, not hype.
Learn more→ AgentsAI agents
Agents with tools, memory and human oversight: support, ops and back office.
Learn more→ ModelsHow we choose a model
Language, code, agents, cost and privacy: the evaluation we run before every deployment, with 30 real prompts per client.
Learn more→
Talk to Slash
Let's put it in production
Tell us your challenge. We reply within 24 business hours with an honest first read: if we can help, we'll say how; if not, we'll say who can.
I reply personally. No endless forms, no canned replies.