Union Alpha Is Now Unbiased Pareto — And the Model Rotation for September 2026

As of September 17, 2026, the stealth endpoint stealth/union-alpha has been revealed. Hitting it now returns a 404 that explicitly redirects to unbiased/pareto.

OpenRouter's catalog lists Pareto as a multimodal composite model with 262,144 tokens of context, currently priced at $2.50/M input / $7.50/M output through OpenRouter.

This isn't a conventional monolithic model. Unbiased describes Pareto as running multiple models against a request and synthesizing/selecting the result.

So the correct model ID to use going forward is:

unbiased/pareto

Not:

stealth/union-alpha

The Practical Rotation (as of September 17, 2026)

For large-repo TypeScript/Svelte/Android work with terminal use, autonomous agent loops, and multi-file edits, here's the rotation I'd actually use:

Primary Daily Driver

DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash)

1.05M context, ~$0.15/M in / $0.60/M out
Released September 10 — new Causal Encoder-Decoder architecture (552B total, 8B/16B active)
Native tool calls, structured outputs, ~104 tok/sec on OpenRouter
Benchmarks: Intelligence Index 39.5, AA long-context 84%, GDPval 56.6%, SciCode 51.9%
Use for: "Inspect this repo, understand the architecture, implement X, run tests, fix failures, don't stop until it works"

Heavy Reasoning / Escalation

GLM 5.3 (z-ai/glm-5.3)

1.31M context, ~$1/M in / $3.41/M out
Artificial Analysis coding score: 74.8 (vs V4 Flash 0731's 69.1, MiMo Pro's 60.2)
Adjustable low/high/max reasoning, reasoning always enabled
Use for: "Why is this architecture failing? Trace this bug across six packages. Design the correct migration rather than merely patching it."

Cheap Bulk Work

DeepSeek V4 Flash 0731 (deepseek/deepseek-v4-flash-0731)

1.31M context, $0.03–0.05/M in / $0.10–0.16/M out — absurdly cheap
AA coding: 69.1, Agentic: 41.7
27 OpenRouter providers
Use for: Analyze 200 files, write tests for 18 components, perform repo-wide migrations, review generated files for regressions
So cheap you can afford to have it review itself twice

Free Experimental Parallel Worker

Unbiased Pareto (unbiased/pareto)

262K context, multimodal composite (runs multiple models + synthesizes)
$2.50/M in / $7.50/M out — but also appears in free lanes
GPQA Diamond: 90.9% — but no public Coding/Agentic Index results yet
~10s P50 latency, ~20 tok/sec
Use for: Free parallel attempt on complicated implementations, then have it critique DeepSeek V4.1 Flash's result

Free Second Opinion / Agent

Big Pickle via OpenCode Zen (opencode/big-pickle)

Stealth model, free, temporary (feedback gathering)
OpenCode doesn't disclose the underlying model — backend may change
Use for: Second opinions, implementation attempts, bug investigations, code review, exploratory refactors
Don't send confidential client repos through it (data may be used to improve the model)

Optional

MiMo V2.5 Pro (xiaomi/mimo-v2.5-pro)

1.05M context, positioned for 1,000+ tool calls
But: OpenRouter endpoint does not accept native tools — awkward for Hermes/OpenCode-style agents
AA coding: 60.2, Agentic: 22.7

One Model I'd Add: GLM 5.3 Flash

It wasn't in the original comparison but deserves a spot:

Coding 71.5, Agentic 51.2, Intelligence 41.9 (Artificial Analysis)
1.31M context
Agentic score higher than V4 Flash 0731's 41.7, while costing far less than full GLM 5.3

The Pareto Set

After looking at current numbers, the practical rotation becomes:

Role	Model	Why
Default / fast agent	DeepSeek V4.1 Flash	Best performance/$ for daily agent loops
Cheap strong reasoning	GLM 5.3 Flash	Higher agentic than V4 Flash 0731, fraction of GLM 5.3 cost
Ultra-cheap bulk	DeepSeek V4 Flash 0731	Costs so little you can double-review everything
Escalation / hardest	GLM 5.3	When you genuinely need deep reasoning
Free parallel worker	Unbiased Pareto	Composite model, free tier available
Free second opinion	Big Pickle (OpenCode Zen)	Unknown strong coding agent, free

This gives you very little reason to burn expensive frontier-model tokens for normal development work.

Quick Reference

PRIMARY
DeepSeek V4.1 Flash
 
HEAVY REASONING / ESCALATION
GLM 5.3
 
CHEAP BULK WORK
DeepSeek V4 Flash 0731
 
FREE EXPERIMENTAL
Unbiased Pareto
 
FREE SECOND OPINION
Big Pickle via OpenCode Zen
 
OPTIONAL
MiMo V2.5 Pro

Or conceptually:

                     ┌─────────────────────┐
                     │ DeepSeek V4.1 Flash │
                     │      DEFAULT        │
                     └──────────┬──────────┘
                                │
                difficult / uncertain result
                                ▼
                     ┌─────────────────────┐
                     │      GLM 5.3        │
                     │   deep reasoning    │
                     └─────────────────────┘
 
 
large / repetitive work ───► DeepSeek V4 Flash 0731
 
free parallel attempt ─────► Unbiased Pareto
 
free critique / experiment ─► Big Pickle