Agnes 2.5 Pro from Singapore Sapiens AI claims 82.7 SWE-bench Verified, 78.7 Multilingual — free OpenAI-compatible API. explainx.ai maps internal evals vs Opus 4.8, GLM 5.2, DeepSeek V4, and what is not open source.
On July 13, 2026, @RoundtableSpace amplified Agnes 2.5 Pro — a free API from Sapiens AI claiming 82.7 on SWE-bench Verified, 78.7 on Multilingual, and SWE Atlas gains that beat GLM 5.2 and DeepSeek V4 Pro on several cuts. The post hit 59.6K views in hours.
The frontier map is no longer US + China only — but read the fine print: these are internal evaluations, the weights are not open, and Opus 4.8 still leads on the hardest bars.
Trails on Verified/Pro/Terminal · leads Multilingual
Singapore frontier?
Product company, not NAIS sovereign training
Who is Agnes AI?
Field
Detail
Company
Sapiens AI (parent) · Agnes AI (product/API)
Entity
Singapore Sapiens Technology PTE
Founded
July 2025
Founder
Bruce Yang — Raffles Institution · NUS AI PhD track · UC Berkeley CS + Applied Math
Funding
~$20M total · $10M Series A (Feb 13, 2026, LOOK FORWARD VCC)
Users
6M+ (Feb 2026, per TechTimes)
ARR
Approaching $20M (Mar 2026)
IPO rumor
Singapore Exchange listing target end of 2026
Agnes positions as world-class multimodal AI for everyone — text, image, video on one OpenAI-compatible gateway. Prior third-party rankings (May 2026): Claw-Eval top 10, Artificial Analysis image top 20, PinchBench top 10 — Agnes claims first Singapore-founded lab at that tier.
Agnes published an INTERNAL EVALUATION comparing Agnes 2.5 Pro, 2.5 Flash, and 2.0 Flash against GLM 5.2, DeepSeek V4 Pro 1.6T, Claude Opus 4.8, and Qwen3.5 397B.
Agnes 2.5 Pro — headline numbers
Benchmark
Agnes 2.5 Pro
Best competitor on chart
Leader
Terminal-Bench 2.1
77.3
Opus 85.0
Opus
SWE-bench Verified
82.7
Opus 87.6 · DeepSeek 80.6
Opus
SWE-bench Pro
61.8
Opus 69.2 · GLM 62.1
Opus
SWE-bench Multilingual
78.7
DeepSeek 75.2 · GLM 73.3
Agnes
SWE Atlas — QnA
40.8
Opus 48.8 · DeepSeek 27.2
Opus
SWE Atlas — RF
42.4
Opus 46.7
Opus
SWE Atlas — TW
38.9
(Qwen 18.5)
Agnes
explainx.ai read: Agnes 2.5 Pro is competitive with DeepSeek V4 and GLM 5.2 on several cuts, not yet Opus-class on Verified/Pro/Terminal — but free changes the economics for agent builders testing harnesses.
Agnes 2.5 Flash — generational jump from 2.0
Benchmark
2.0 Flash
2.5 Flash
Delta
Terminal-Bench 2.1
52.6
62.3
+9.7
SWE-bench Verified
72.4
75.8
+3.4
SWE-bench Pro
49.6
50.4
+0.8
SWE-bench Multilingual
67.3
69.1
+1.8
SWE Atlas QnA
15.8
36.5
+20.7
SWE Atlas RF
11.4
29.5
+18.1
SWE Atlas TW
13.5
27.5
+14.0
SWE Atlas is a newer agentic coding suite — Ornith-1.0 docs describe QnA/RF/TW harnesses with mini SWE agents. Agnes's 2.0 → 2.5 Flash jump on Atlas is the story Mario Nawfal highlighted — if reproducible, it signals real agent-training progress, not just Verified saturation.
Critical caveat — internal eval
The chart is labeled INTERNAL EVALUATION. Unless Agnes publishes:
When Opus wins: Hardest autonomous coding, Fable-class depth, production where 3–5 pt Verified gap matters.
vs GLM 5.2 (open weights)
Dimension
Agnes 2.5 Pro
GLM 5.2
SWE-bench Multilingual
78.7
73.3
Terminal-Bench 2.1
77.3
81.0
SWE-bench Pro
61.8
62.1
License
API-only
MIT · self-host
Cost
Free tier RPM limits
GPU + hosting
@preferredev_ on X: "probably built on another open source model like GLM, Deepseek, or Kimi" — unverified. Agnes markets indigenous multimodal training; no public architecture paper yet.
vs DeepSeek V4 Pro
Dimension
Agnes 2.5 Pro
DeepSeek V4 Pro
SWE-bench Verified
82.7
80.6
SWE-bench Multilingual
78.7
75.2
Terminal-Bench 2.1
77.3
64.0
SWE Atlas QnA
40.8
27.2
Access
Free Singapore API
DeepSeek API / weights TBD
Agnes leads on terminal + Atlas in this chart; DeepSeek still competitive on Verified.
Singapore and the frontier narrative
Our Singapore AI landscape (June 2026) framed the city-state as trusted hub + governance, not indigenous GPT-class training. Agnes complicates that story:
Not the same as: US frontier labs (OpenAI, Anthropic) or China sovereign stacks (Zhipu, DeepSeek, Qwen). Agnes is a venture-backed API company — closer to regional product innovation than NAIS compute sovereignty.
Geopolitical angle from X:@UsmanAnzaar — "More countries releasing models = larger attack surface." Free multimodal APIs lower builder friction; they also expand prompt-injection and data-exfil surfaces for agent tools. Same threat model as Grok build repo secrets and DCG for coding agents.
What developers should do now
Register and run 50 real tickets from your backlog — not leaderboard faith
Pair with local guardrails — DCG if agents get shell access
Agent loop prompt (copy-paste)
snippet
You are a coding agent using Agnes 2.5 Pro.
Task: {one real GitHub issue from our repo}
Rules:
- Read failing test output first
- Propose minimal patch
- Run tests before claiming done
- Stop if rate-limited (429) and report
Benchmark values reflect Agnes internal evaluation as circulated July 13, 2026. Reproduce on your own tasks before changing production defaults — vendor charts are starting points, not warranties.
