Dense specification format. Machine-readable. Maximum signal density.
Metadata
---
spec: openrouter-agent-models-v1
date: "2026-03-14"
models_described: 2
data_source: "openrouter_api /api/v1/models"
query_date: "2026-03-14"
model_creation: "~2026-03-09"
status: "alpha"
---
Model Specs
[hunter_alpha]
id: "openrouter/hunter-alpha"
canonical_slug: "openrouter/hunter-alpha"
name: "Hunter Alpha"
namespace: "openrouter"
provider: "undisclosed"
status: "alpha"
parameters: 1_000_000_000_000 # 1 trillion
context_window: 1_048_576 # 1M tokens
max_completion_tokens: 32_000
modality: "text->text"
input_modalities: ["text"]
output_modalities: ["text"]
tokenizer: "Other"
pricing: {
prompt: 0,
completion: 0,
request: 0,
image: 0,
web_search: 0,
internal_reasoning: 0
}
default_temperature: 1.0
default_top_p: 0.95
is_moderated: false
data_logging: true # all prompts + completions logged
use_case: "long-horizon planning, complex reasoning, multi-step execution"
description_excerpt: "1T param + 1M context frontier intelligence model built for agentic use"
[healer_alpha]
id: "openrouter/healer-alpha"
canonical_slug: "openrouter/healer-alpha"
name: "Healer Alpha"
namespace: "openrouter"
provider: "undisclosed"
status: "alpha"
parameters: "undisclosed"
context_window: 262_144 # 256K tokens
max_completion_tokens: 32_000
modality: "text+image+audio+video->text"
input_modalities: ["text", "image", "audio", "video"]
output_modalities: ["text"]
tokenizer: "Other"
pricing: {
prompt: 0,
completion: 0,
request: 0,
image: 0,
web_search: 0,
internal_reasoning: 0
}
default_temperature: 1.0
default_top_p: 0.95
is_moderated: false
data_logging: true # all prompts + completions logged
use_case: "multi-modal perception, cross-modal reasoning, real-world action"
description_excerpt: "frontier omni-modal model with vision, hearing, reasoning, and action capabilities"
Supported Parameters
[api_parameters]
# Both models share identical parameter support
supported: [
"frequency_penalty",
"include_reasoning", # chain-of-thought exposure
"max_tokens",
"presence_penalty",
"reasoning", # enable/disable reasoning
"response_format", # JSON / structured output
"stop", # stop sequences
"temperature",
"tool_choice", # function selection control
"tools", # function calling
"top_p"
]
not_supported: [
"logprobs", # no distributional introspection
"top_logprobs",
"seed", # no deterministic reproducibility
"structured_outputs", # use response_format instead
"top_k"
]
# Omission analysis: missing logprobs + seed suggests
# optimization for execution confidence over sampling
# exploration. Consistent with agent use where models
# commit to actions rather than hedge across distributions.
Architectural Split
[design_pattern]
pattern: "brain_perception_split"
analogy: "robotics_high_low_level_planning"
[hunter_alpha.role]
function: "deep_reasoning_engine"
optimizes_for: [
"long_horizon_planning",
"multi_step_execution_chains",
"instruction_fidelity_over_time",
"complex_reasoning_depth"
]
architecture: "large_dense_1T_text"
context_budget: 1_048_576
[healer_alpha.role]
function: "perception_action_layer"
optimizes_for: [
"multi_modal_understanding",
"cross_modal_reasoning",
"real_world_sensing",
"environment_awareness"
]
architecture: "omni_modal_efficient"
context_budget: 262_144
parameter_count: "undisclosed" # likely smaller than Hunter
[composition_pattern]
method: "route_by_task_type"
strategy_to_hunter: "planning, reasoning, code, long chains"
strategy_to_healer: "images, audio, video, real-time perception"
framework_compatibility: ["openclaw", "langchain", "crewai", "autogen"]
Comparison to Existing Models
[market_positioning]
[vs_general_purpose]
gpt_4o: { purpose: "general", context: 128_000, modalities: "text+image", pricing: "paid" }
claude_opus: { purpose: "general", context: 200_000, modalities: "text+image", pricing: "paid" }
gemini_3_pro: { purpose: "general", context: 1_000_000, modalities: "text+image+audio+video", pricing: "freemium" }
[vs_purpose_built]
hunter_alpha: { purpose: "agent_reasoning", context: 1_048_576, modalities: "text", pricing: "free_alpha" }
healer_alpha: { purpose: "agent_perception", context: 262_144, modalities: "text+img+audio+video", pricing: "free_alpha" }
[key_differentiators]
hunter: "only 1T purpose-built agent model on OpenRouter"
healer: "only free omni-modal agent model on OpenRouter"
shared: "both free, both tool-native, both alpha (data logged)"
Agent Framework Impact
[scaffold_reduction_analysis]
[current_agent_patterns_general_models]
prompt_scaffolding: "system prompts to simulate agent behavior"
context_management: "compression + summarization to fit limited windows"
tool_reliability: "retry loops for non-agent-optimized models"
instruction_drift: "degradation over long execution chains"
[purpose_built_model_hypothesis]
premise: "agent models reduce scaffolding burden"
if_true: [
"less prompt engineering for agent behavior",
"fewer context compression strategies needed",
"higher first-attempt tool call reliability",
"sustained instruction fidelity over long chains"
]
if_false: [
"alpha models may underperform established generalists",
"no benchmarks to validate claims",
"free tier temporary; pricing unknown"
]
verification: "requires agent benchmark evaluation (SWE-Bench, AgentBench)"
Open Questions
[unknowns]
benchmarks: "none published — SWE-Bench, AgentBench, TerminalBench scores missing"
trainer: "OpenRouter is a router; actual model provider undisclosed"
training_method: "RLHF on agent trajectories? Multi-turn pretraining? Architecture?"
latency: "1T params = expensive serving; real-world agent loop latency unknown"
availability_timeline: "alpha duration + production path undefined"
post_alpha_pricing: "will free tier persist? competitive with Claude/GPT pricing?"
healer_params: "undisclosed — likely smaller than Hunter given multi-modal overhead"
model_stability: "alpha models may change significantly before production"
Assessment
[verdict]
directional_signal: "strong — purpose-built agent models are a real category"
empirical_evidence: "weak — alpha status, no benchmarks, all claims unverified"
practical_value: "high for prototyping — free access, zero barrier to entry"
production_readiness: "not yet — data logging, no SLA, alpha instability"
recommended_action: "experiment with non-sensitive agent workloads; monitor for beta"
[risk_matrix]
data_exposure: "HIGH — all prompts/completions logged"
model_stability: "MEDIUM — alpha, subject to change"
vendor_lock_in: "LOW — OpenRouter standard API, portable"
cost_risk: "LOW — free during alpha; pricing TBD post-alpha"
References
openrouter_models_api: "https://openrouter.ai/api/v1/models"
hunter_alpha_page: "https://openrouter.ai/openrouter/hunter-alpha"
healer_alpha_page: "https://openrouter.ai/openrouter/healer-alpha"
openclaw: "https://openclaw.ai"
data_query_timestamp: "2026-03-14T04:30:00Z"
hunter_created: 1773260671 # unix timestamp
healer_created: 1773260668 # unix timestamp