Browser's /mythos speaks OpenAI-compatible chat-completions — two env vars switch backends.
Updated · 2026-05-22·~8 min read·9 providers covered
1. OpenAI closed
The default for /mythos.
# in .envOPENAI_API_KEY=sk-...# OPENAI_BASE_URL=https://api.openai.com/v1 # default, can omit# In /mythos per-user LLM config: provider="openai", model="gpt-4o-mini"
Best at
Default, balanced
Cost / mythos run
~$0.30-1.50
Latency
~250ms/token TTFT
2. Anthropic API closed
Claude Opus 4.7 / Sonnet 4.6 direct. Mythos Preview is NOT generally available.
# in .envANTHROPIC_API_KEY=sk-ant-...# /mythos LLM config: provider="anthropic", model="claude-sonnet-4-6"# or model="claude-opus-4-7" for the heavyweight
Best at
Adversarial validation
Cost / mythos run
~$1.50-7.00
Latency
~400ms/token TTFT
3. AWS Bedrock cloud
Same Claude routed through your AWS account. Mythos via allowlist only.
# 1. Set AWS creds via your usual chain (env, profile, IAM role)AWS_REGION=us-east-1# Mythos preview regionAWS_PROFILE=your-profile# 2. Use Bedrock Converse API; /mythos shims to OpenAI shape internally# /mythos config: provider="bedrock"# model="anthropic.claude-sonnet-4-6-v1:0"
Best at
Enterprise compliance
Cost / mythos run
~$1.50-7.00 (same as direct)
Bonus
PrivateLink · BAA · audit logs
4. Google Cloud Vertex AI cloud
Claude Mythos Preview is in Vertex AI Private Preview (April 2026).
# 1. GCP service account JSON or workload identityGOOGLE_APPLICATION_CREDENTIALS=/path/to/sa.jsonGOOGLE_CLOUD_PROJECT=your-project-idGOOGLE_CLOUD_REGION=us-east5# Mythos Preview region# /mythos config: provider="vertex"# model="claude-mythos-preview@..." (when allowlisted)# fallback model="claude-opus-4-7@20260408"
Best at
GCP-shop, multi-model
Cost / mythos run
~$1.50-7.00
Bonus
Vertex eval suites · model garden
5. Groq closed inferenceopen weights
Groq runs open-weight Llama 3.3 70B at 5-10× GPU speed.
# OpenAI-compatible — same env vars, different base URLOPENAI_BASE_URL=https://api.groq.com/openai/v1OPENAI_API_KEY=gsk_...# /mythos model: "llama-3.3-70b-versatile" or "deepseek-r1-distill-llama-70b"
Best at
Speed-critical workflows
Cost / mythos run
~$0.05-0.30 (free tier OK for testing)
Latency
~30ms/token TTFT
6. Mistral La Plateforme cloudopen weights
EU-based, GDPR-aligned, data-residency in France.
# OpenAI-compatible endpointOPENAI_BASE_URL=https://api.mistral.ai/v1OPENAI_API_KEY=your-mistral-key# /mythos model: "mistral-large-latest" or "open-mistral-nemo"
Best at
EU data-residency
Cost / mythos run
~$0.20-0.80
Latency
~200ms/token TTFT
7. Ollama (self-host) open weights
Air-gapped path. Self-host on your GPU box.
# Server (host with GPU):
$ ollama pull llama3.3:70b
$ ollama serve # exposes :11434# Browser config (.env):OPENAI_BASE_URL=http://your-gpu-host:11434/v1OPENAI_API_KEY=ollama# any non-empty value# /mythos model: "llama3.3:70b"
Best at
Air-gapped, classified, sovereign
Cost / mythos run
$0.30/h amortised GPU
Latency
depends on hardware
8. OpenMythos (community reconstruction) open weights · MIT
kyegomez/OpenMythos — community reconstruction of the Mythos architecture, MIT-licensed, 13k+ stars.
# 1. Install + run on your GPU server (requires ~80GB VRAM for 70B class)
$ pip install open-mythos vllm
$ vllm serve open-mythos/mythos-50b --host 0.0.0.0 --port 8000
# 2. Browser .envOPENAI_BASE_URL=http://your-vllm-host:8000/v1OPENAI_API_KEY=EMPTY# vLLM accepts any token
Best at
Research, architecture validation
Cost
Hardware-only
Caveat
Not the real Mythos · ~50-65% quality
9. Together AI closed inferenceopen weights
Together AI hosts a wide bench of open-weight models at competitive prices.
OPENAI_BASE_URL=https://api.together.xyz/v1OPENAI_API_KEY=your-together-key# /mythos model examples:# meta-llama/Llama-3.3-70B-Instruct-Turbo# deepseek-ai/DeepSeek-R1-Distill-Llama-70B# Qwen/Qwen2.5-72B-Instruct-Turbo
Cost matrix — per /mythos run (50-100 LLM calls)
Estimated all-in cost per full /mythos run.
Provider
Model
In $/1M
Out $/1M
/run estimate
Groq
Llama 3.3 70B
$0.59
$0.79
~$0.05–0.30
OpenAI
GPT-4o-mini
$0.15
$0.60
~$0.30–1.50
Together AI
Llama 3.3 70B Turbo
$0.88
$0.88
~$0.20–1.00
Mistral
Mistral Large 2
$2.00
$6.00
~$0.20–0.80
Anthropic
Sonnet 4.6
$3.00
$15.00
~$1.50–7.00
AWS Bedrock
Claude Sonnet 4.6
$3.00
$15.00
~$1.50–7.00
Vertex AI
Claude Sonnet 4.6
$3.00
$15.00
~$1.50–7.00
Anthropic
Opus 4.7
$15.00
$75.00
~$5.00–25.00
Anthropic
Mythos Preview (Glasswing)
$25.00
$125.00
~$10.00–60.00
Ollama / vLLM
Llama 3.3 70B (self-host)
$0
$0
hardware-only ~$0.30/h
Decision tree — which one for which constraint?
⌗ BUDGET-CONSTRAINED
Groq · OpenAI · Together AI
Groq has the cheapest per-run cost and fastest latency.
⌗ COMPLIANCE / DATA-RESIDENCY
AWS Bedrock · Vertex AI · Mistral · self-host
Route through your existing cloud vendor or self-host.
Groq's LPU silicon — 8× faster than typical GPU inference.
Switching backends takes 2 env vars and a model name.
Every recipe on this page is tested with Browser's /mythos harness. Pick one that matches your budget, compliance, and latency — point /mythos at it, and run.
Browser