best new entrant
Hark
97.7 OM2W · founded 2026
A compelling browser-use score paired with unusually low stated token pricing. Validate reliability on long, messy workflows.
Six young AI startups now post narrow, measurable wins over established rivals
A scan of recent model launches found six companies founded since 2023 with explicit head-to-head results against established AI products. Liquid AI shows the largest reported edge—3.7× faster in one tool-use test—but all six comparisons are company-reported, so these are scouting signals, not settled market leadership.
6
credible leads to investigate
2023–26
founding years
3.7×
largest reported speed ratio
0
fully independent tests found
Reported uplift over the named rival
Percent improvement; different benchmarks are not directly comparable
blue = company-reported edge
Liquid AI
+270%
Thinking Machines
+66%
Fish Audio
+50%
Poolside
+9.7%
Hark
+5.3%
Perceptron
+0.7%
0%135%270%
What stands out
1
Hark is the newest and most price-aggressive. Its 2026 browser agent scored 97.7 against GPT‑5.4’s 92.8 while listing token prices far below the named frontier model. venturebeat.com
2
Thinking Machines combines lower latency with higher interaction quality. Its voice/video model reported 0.40-second turn taking versus Gemini’s 0.57 and OpenAI’s 1.18 seconds, plus a 77.8 quality score versus OpenAI’s 46.8. venturebeat.com
3
Liquid AI’s edge is deployment efficiency. Its compact on-device model finished a 35-tool-call workload 3.7× faster than DeepSeek‑V4‑Flash. venturebeat.com
4
The evidence is promising but early. Every shortlisted result originates with the vendor; reproduce the tests on your own workload before treating a benchmark win as durable superiority.
Three strongest scouting targets
best real-time interaction
Thinking Machines
0.40 sec · founded 2025
The cleanest combined latency-and-quality claim in the set. Access and repeatability remain the key diligence questions.
best edge efficiency
Liquid AI
3.7× faster · founded 2023
Most interesting where cloud dependency, latency, or device constraints matter more than raw general intelligence.
Full shortlist
| Startup | Founded | Product | Compared with | Reported result | Source |
|---|---|---|---|---|---|
| Hark | 2026 | Handoff | GPT 5.4; Claude Opus 4.8 | 97.7 vs 92.8 / 84.1 on OM2W; sharply lower token price | venturebeat.com |
| Thinking Machines | 2025 | TML-Interaction-Small | Gemini live; GPT realtime | 0.40s latency; 77.8 interaction quality | venturebeat.com |
| Perceptron | 2024 | Mk1 | Robotics‑ER 1.5; Q3.5‑27B | 85.1 vs 78.4 / about 84.5 | venturebeat.com |
| Fish Audio | 2023 | S2 Pro | ElevenLabs V3 | 60% vs 40% listener preference | fish.audio |
| Liquid AI | 2023 | LFM2.5‑2.6B | DeepSeek‑V4‑Flash | 3.7× faster on a 35-call tool workload | venturebeat.com |
| Poolside | 2023 | Laguna S 2.1 | DeepSeek‑V4‑Pro‑Max | 70.2% vs 64.0% on Terminal‑Bench 2.1 | venturebeat.com |
Method: web scan performed 22 August 2026 across recent launch coverage, vendor posts, benchmark comparisons and startup directories; six companies founded in 2023 or later survived the requirement for a named rival and a quantified head-to-head result. “Uplift” measures relative change on each source’s own stated metric and is not a cross-domain ranking.