Skip to content

Latest commit

 

History

History
216 lines (155 loc) · 11.4 KB

File metadata and controls

216 lines (155 loc) · 11.4 KB

SOTA vs fak — SWE-bench Verified Comparison

Date: 2026-06-20 Status: ✅ Infrastructure complete, awaiting real agent integration

Executive Summary

Aspect SOTA Leaders fak (Current) fak (Potential)
Resolve Rate 72-77% Not measured TBD (needs real agent)
Cost Efficiency Prefix prompt-cached per session + cross-worker/session reuse (~1.1–1.2× at 2-4 workers) Scales with shared-prefix fraction & workers
Cross-worker Reuse None (single-agent / single-instance) 1.13-1.22x at 2-4 workers Scales with workers
Turn Tax Already reused via prompt cache KV persistence (≈ parity) Eliminated
Infrastructure mini-SWE-agent (100-line ReAct) Production-grade gateway Multi-model, multi-worker

Bottom line: the SOTA leaders run on API models that already prompt-cache the shared prefix across turns, so the honest infra comparison is not a 20×. fak's incremental win is cross-worker / cross-session prefix sharing — ~1.1–1.2× at 2-4 workers, growing with the shared-prefix fraction — plus addressable eviction and a default-deny safety floor those stacks don't offer. (A 20–24× number exists only against a naive re-prefill-every-turn loop — a worst case no SOTA run ships — and is never the comparison.) With a competitive model (Claude/GPT class), fak could match SOTA resolve rates at a modestly lower and more scalable infra cost.


Current SOTA Leaders (SWE-bench Verified)

Data from swebench.com as of February 2026.

Top 10 Models (mini-SWE-agent v2 harness)

Rank Model % Resolved Avg Cost $ Org Date
1 Claude 4.5 Opus (high reasoning) 76.80% $0.75 Anthropic 2026-02-17
2 Gemini 3 Flash (high reasoning) 75.80% $0.36 Google 2026-02-17
3 MiniMax M2.5 (high reasoning) 75.80% $0.07 MiniMax 2026-02-17
4 Claude Opus 4.6 75.60% $0.55 Anthropic 2026-02-17
5 GPT-5-2 Codex 72.80% $0.45 OpenAI 2026-02-19
6 GLM-5 (high reasoning) 72.80% $0.53 Z-AI 2026-02-17
7 GPT-5-2 (high reasoning) 72.80% $0.47 OpenAI 2026-02-17
8 Claude 4.5 Sonnet (high reasoning) 71.40% $0.66 Anthropic 2026-02-17
9 Kimi K2.5 (high reasoning) 70.80% $0.15 Moonshot 2026-02-17
10 DeepSeek V3.2 (high reasoning) 70.00% $0.45 DeepSeek 2026-02-17

Common Infrastructure

All SOTA methods use the same evaluation harness:

  • mini-SWE-agent — 100-line Python ReAct loop
  • Bash-only environment — No tools, just LM + shell
  • Temperature 0.0 — Deterministic (for release 1.x)
  • KV reuse happens at the provider — the harness re-sends the transcript each turn, but the API models it calls (Claude / GPT / Gemini) prompt-cache the shared prefix (system prompt + tool schemas + earlier turns), read back at ~0.1× cost. So SOTA runs do not re-prefill the shared prefix from scratch every turn; the naive "re-prefill everything" loop is a worst case nobody runs, not the SOTA baseline.

This means the differences in resolve rate come from:

  1. Model capability (not infrastructure)
  2. Prompt engineering (within mini-SWE-agent)
  3. Reasoning mode ("high reasoning" flag)

fak's Value Proposition

Infrastructure Advantages (Measured)

From the smoke test on 5 instances (difficulty-derived geometry):

Prefill Work-Elimination (A/C Ratio)

Workers A (Naive) B (Per-Agent KV) C (fak Fused) A/C B/C
1 1.04M tokens 52.9K tokens 52.9K tokens 19.7x 1.00x
2 2.09M tokens 105.8K tokens 93.3K tokens 22.4x 1.13x
4 4.17M tokens 211.6K tokens 174.1K tokens 24.0x 1.22x

Interpretation:

  • A/C (Net Work-Elimination): fak reduces 95%+ of prefill work vs naive re-prefetch-every-turn
  • B/C (Cross-Worker Reuse): Shared prefix gives 1.22x benefit at 4 workers (the value stack)
  • A/B (Turn Tax): 19.7x — re-prefill vs KV persistence, worker-independent

Which of these is the SOTA comparison? Only B/C. A/C and A/B are against the naive re-prefill loop — a worst case no SOTA run ships. A real SOTA setup (provider prompt caching / per-instance prefix caching) already eliminates the same turn-tax, so fak's incremental infra win over it is the cross-worker / cross-session B/C reuse (~1.1–1.2× at 4 workers, larger as the shared-prefix fraction grows), not the 20–24× floor.

What This Means in Practice

For a full SWE-bench Verified run (500 instances, ~20 turns median):

Approach Prefill Tokens (est.) vs the realistic SOTA
Naive (re-prefill every turn) — worst-case floor, not SOTA ~500M tokens
Per-Agent KV / provider prompt cache — the SOTA baseline ~25M tokens baseline
fak (shared prefix across workers/sessions) ~21M tokens ~1.1–1.2× at small scale, rising with shared-prefix fraction

On SWE-bench's ~500 instances:

  • SOTA methods already reuse the shared prefix (provider prompt caching / per-instance prefix caching) — they do not pay the full prefill tax every turn.
  • fak adds cross-worker and cross-session sharing on top, so the same system prompt + tool schemas are shared across instances, not just across turns within one.
  • Net vs the realistic SOTA: a ~1.1–1.2× infra reduction at a few workers, growing with the shared-prefix fractionnot the 20–24× that only appears against the naive floor.

Comparative Analysis

What SOTA Does Better

  1. Model Integration — Direct access to Claude/GPT/DeepSeek APIs
  2. Proven Resolve Rates — 70-77% on the official leaderboard
  3. Simplicity — 100-line Python agent, easy to run
  4. Ecosystem — Extensive tooling, papers, community

What fak Does Better

  1. Infrastructure Efficiency — cross-worker / cross-session prefix sharing (~1.1–1.2× at a few workers over a SOTA stack that already prompt-caches; the 20–24× is only vs a naive re-prefill floor)
  2. Scalability — Cross-worker reuse (1.22x at 4 workers), rising with the shared-prefix fraction
  3. Production Readiness — Gateway architecture, multi-model support
  4. Session Persistence — KV cache reuse across turns
  5. Adjudication — In-process vs spawn-per-hook (measurable via fak bench)

The Gap

Aspect SOTA fak Path to Parity
Resolve Rate 70-77% Not measured Integrate real model + agent
Model Access Direct API Placeholder Add Claude/GPT endpoint
Agent Logic ReAct loop Placeholder Port mini-SWE-agent logic

Key insight: fak's infrastructure advantages are orthogonal to model capability. With the same model (e.g., Claude 4.5 Opus), fak could achieve ~77% resolve rate at a modestly lower, fleet-scaling infrastructure cost — cross-worker/session prefix sharing on top of the prompt caching the SOTA already does (not the 20–24× that only holds vs a naive re-prefill loop).


Projections

Cost Comparison (Hypothetical Full Run)

Assuming:

  • 500 instances (SWE-bench Verified)
  • Median 20 turns per instance
  • 200 tokens/turn (assistant) + 400 tokens/turn (tool results)
  • Claude 4.5 Opus at $3/M input, $15/M output
Approach Input Tokens (billable) Output Tokens Est Cost
Naive, no reuse — worst-case floor; no SOTA run pays this ~2.0B ~1.0B ~$21,000
Realistic SOTA (provider prompt cache / per-instance prefix cache) far lower — the shared prefix is cache-read at ~0.1× ~1.0B a small fraction of the floor
fak (+ cross-worker / cross-session sharing) lower still by the shared-prefix fraction ~1.0B ~1.1–1.2× below the SOTA at a few workers

Note: the "$21,000 → ~93% savings" figure is against the naive floor, which no SOTA run pays. Output tokens are identical (same generated code). Against the realistic SOTA — which already prompt-caches the shared prefix — fak's incremental win is sharing the prefix across workers and sessions (the B/C delta), not the 20× turn-tax a tuned engine already eliminates.

Performance Projection

If fak integrated a SOTA model:

Metric SOTA fak (with SOTA model)
Resolve Rate 70-77% 70-77% (same)
Infrastructure Cost Already prompt-cached per session ~1.1–1.2× lower at a few workers (cross-worker/session sharing); grows with shared-prefix fraction
Latency Baseline Lower (shared prefix prefilled once across workers)

The win: same quality; a cross-worker/session infra reduction that scales with fleet size — not the 20× that only holds against a naive re-prefill loop.


Next Steps

Immediate (Ship Real Agent)

  1. Fleet Gateway Integration

    • Connect fak swebench run --agent fleet to real gateway
    • Implement actual agent loop (ReAct or similar)
    • Test with Claude 4.5 Opus endpoint
  2. Model Access

    • Add Anthropic API key configuration
    • Implement proper tool calling (not placeholder)
    • Add trajectory capture for analysis
  3. Resolve Rate Measurement

    • Run full smoke test with real model
    • Compare predictions to gold patches
    • Publish first fak resolve rate

Medium (Competitive Evaluation)

  1. Head-to-Head with SOTA

    • Run Claude 4.5 Opus via fak infrastructure
    • Compare resolve rate to mini-SWE-agent baseline
    • Quantify cost savings
  2. Full Dataset Run

    • Scale from 5 instances to 500
    • Measure infrastructure cost at scale
    • Publish fak leaderboard entry
  3. Advanced Features

    • Multi-worker comparison (1, 2, 4, 8 workers)
    • Real-time adjudication measurement
    • Prometheus metrics integration

Long Term (Research Directions)

  1. Model-Specific Optimization

    • Fine-tune prompts for each model class
    • Implement "high reasoning" mode equivalents
    • Ablate value stack benefits per model
  2. Beyond SWE-bench

    • WebBench (frontier web tasks)
    • Custom enterprise benchmarks
    • Multi-modal code tasks

Conclusion

fak's infrastructure is ready for SOTA competition. Its incremental infra win over the realistic SOTA — API models and engines that already prompt-cache the shared prefix — is cross-worker / cross-session prefix sharing (~1.1–1.2× at a few workers, growing with the shared-prefix fraction) plus addressable eviction and a safety floor; the 20–24× figure holds only against a naive re-prefill loop nobody ships. With proper model integration, fak could match the 70-77% resolve rate of current leaders at a modestly lower and more scalable infrastructure cost.

The blocking factor is not infrastructure — it's agent implementation. The placeholder runners need real model access and proper ReAct loop logic. Once that lands, fak can publish competitive resolve rates while demonstrating superior cost efficiency.

Sources: