Latent Operator

Stop wasting time on overhyped AI models. We expose what works, what fails, and what’s not worth your money. Real tests. Brutal comparisons.

[Latent Operator @LatentOperatorX]

DeepSeek Harness VS Opencode Harness VS GPT-6.1 Terra on an A380 takeoff ✈️🤯

Blind AI judge, same prompt:
DSH: 9.0/10 · 69m · ~$0.50
Terra: 7.9/10 · 11m · ~$0.22
OpenCode Harness: 6.9/10 · 11m · ~$0.05

DeepSeek runs: V4.1 Flash via OpenCode Go.

Worth it? Follow @LatentOperatorX for more AI battles. ⚔️

[Latent Operator @LatentOperatorX]

Most bad AI outputs are not model failures. They are missing context. If the model had to guess, it guessed wrong. Give it the constraints you would give a new hire.

[Latent Operator @LatentOperatorX]

@AAksola @opencode I’ve seen others reporting lower tokens-per-second rates than me on Command Code, but yeah, those are my results. Maybe it depends on the location, although we should all be getting the same level of service regardless of where we are in the world.

[Latent Operator @LatentOperatorX]

📊 48 HOURS LATER: DeepSeek V4.1 Flash on @opencode, here’s what the data shows.

Two days ago, I reported serious performance issues with DeepSeek V4.1 Flash on OpenCode Go.

The response from the community was incredible, and the OpenCode team acknowledged the issue.

Since then, I’ve moved my benchmarks to @CommandCodeAI while continuing to monitor OpenCode Go.

Here are my latest measurements:

⚡ Command Code: 168 tokens/s median (14 benchmark runs)

🐢 OpenCode Go: 69 tokens/s median (66 probes, 12-hour rolling trend)

That’s a 2.4× difference in observed output speed using the same model.

But here’s the encouraging part. 📈

OpenCode Go’s daily median is improving:

Nov 7: 66 tok/s
Nov 8: 69 tok/s
Nov 9: 95 tok/s

Progress is happening. The gap hasn’t closed yet, but the trend is encouraging.

Important context: these measurements use different workloads, so this isn’t a perfectly controlled head-to-head comparison. Output speed also doesn’t tell the whole story about model quality.

Credit to @fanjiewang and the OpenCode team for their transparency and responsiveness.

Frank has also hinted that something better is coming. 👀

I’ll keep monitoring performance, stability and real benchmark results as improvements roll out.

No sponsorship. Just independent testing and transparent data.

Good or bad, I’ll share what the numbers show.

@LatentOperatorX ⚔️

[Latent Operator @LatentOperatorX]

What happened to DeepSeek V4.1 Flash on @opencode? 💀

For the past 3 days, it’s been almost unusable. EXTREMELY slow, and struggling with even basic tasks.

I can’t even run my AI benchmarks anymore.

Am I the only one experiencing this?

@thdxr, what’s going on?

[Latent Operator @LatentOperatorX]

@Issssyf @opencode I recorded 7 overload errors on Nov 7, but none in my more recent probes.

That’s definitely encouraging, although it’s too early to say the issue is fully resolved.

[Latent Operator @LatentOperatorX]

🦊 DeepSeek Harness vs NO Harness..
Is the quality worth the wait?! 🤯

Same DeepSeek V4.1 Flash model.
Same prompt. Same MAX reasoning effort.

���️ Challenge: Create a voxel-style red fox in a snowy forest.

🟢 DeepSeek Harness
⭐ 8.6/10
⏱️ 171m 49s
💰 ~$0.69

🔵 No Harness (Direct API)
⭐ 6.1/10
⏱️ 9m 05s
💰 ~$0.07

💀 The difference is INSANE:

🏆 DeepSeek Harness scored 2.5 points higher.

⚡ But it took 19× longer and cost almost 10× more!

Same model. Two completely different experiences.

One run per setup. AI-judged scores.

Be honest: Would you wait almost 3 HOURS for the better result? 👀

👇 Which version would YOU choose?

Follow @LatentOperatorX for more AI battles. ⚔️

[Latent Operator @LatentOperatorX]

⚔️ DeepSeek Harness vs Codex Harness 👉 CRASH TEST! 🚗💥

Same DeepSeek V4.1 Flash model. Same prompt. Different harnesses and reasoning effort.

I challenged both to simulate a small car crashing into TWO concrete mixer trucks and a BATTLE TANK. 💀

🔵 Codex Harness (High Effort)
⭐ 4.0/10
⏱️ 48m 38s
💰 ~$0.24

🟢 DeepSeek Harness (Max Effort)
⭐ 6.5/10
⏱️ 120m
💰 ~$0.46

🤯 DeepSeek Harness scored 2.5 points higher… but took 2.5× longer and cost almost TWICE as much!

Scores are AI-judged. One run per setup.

Would you sacrifice speed and cost for better quality? 👀

👇 Which result impressed you more?

Follow @LatentOperatorX for more AI battles. ⚔️

[Latent Operator @LatentOperatorX]

🏆 THE HARNESS CUP — ROUND 3/4

Codex Harness vs Claude Code Harness vs DeepSeek Harness! ⚔️🐎

Same Muse Spark 1.3 model. Same prompt. Same high reasoning effort.

🐎 Challenge: Build a 3D horse show-jumping animation in Three.js!

🔵 Codex Harness: 6.3/10 · 15m 16s · $1.05

🟠 Claude Code Harness: 5.7/10 · 21m 37s · $1.18

🟢 DeepSeek Harness: 5.2/10 · 10m 08s · $1.21

🤔 The AI judge picked Codex Harness… but honestly, I’m not fully convinced.

I’m genuinely struggling to decide which harness deserves to win this round.

I need YOUR verdict. 👀

👇 Codex Harness, Claude Code Harness or DeepSeek Harness?

🏆 Most votes takes the round!

Follow @LatentOperatorX for more AI battles. ⚔️

#AI #CodingAgents #Benchmark

[Latent Operator @LatentOperatorX]

DeepSeek V4.1 Flash + the Codex harness is a seriously underrated duo. It COOKED this crash test. 🔥

I challenged it to simulate a small car crashing into TWO concrete mixer trucks… and a BATTLE TANK. 🚗💥

Watch the survival odds plummet. 💀

But here’s the real question: Does the physics actually hold up? 👀

DeepSeek Harness version coming in the replies ���

Follow @LatentOperatorX for more AI battles. ⚔️

[Latent Operator @LatentOperatorX]

@laiyijulia35276 Next one will be better 😌

[Latent Operator @LatentOperatorX]

DeepSeek Harness vs Codex Harness ⚔️
Who runs DeepSeek V4.1 Flash better?
Same model. Same prompt. Same high reasoning effort.Only the harness changes.

Challenge: Build a water bomber animation in Three.js. 🛫🚿🧑🏽‍🚒

🔵 Codex Harness
• Time: 65m 48s
• Cost: ~$0.25

🟢 DeepSeek Harness
• Time: 120m
• Cost: ~$0.30

One run per setup.

Same AI. Different harness. Different results.

Which harness would YOU choose? 👇

Follow @LatentOperatorX for more AI battles. ⚔️

#AI #DeepSeek #Codex

[Latent Operator @LatentOperatorX]

@hany99dev To frustrate them.

[Latent Operator @LatentOperatorX]

@Sinan_ADCode No, it’s mostly due to agent loop overhead and token efficiency. The DeepSeek harness is significantly slow.

[Latent Operator @LatentOperatorX]

DeepSeek V4.1 Flash Fast with a HARNESS: +5.3 POINTS?! 🤯🏎️

The F1 pit stop test! 🛞

Same DeepSeek V4.1 Flash Fast. Same provider (Command Code). Same prompt. HIGH effort.

🧠 DeepSeek Harness: 7.9/10 · 73m · ~$0.93 · 264k tokens
⚡ No harness: 2.6/10 · 8m · $0.05 · 86k tokens

Higher AI-judged score, ~19× the cost. Worth it?

Follow @LatentOperatorX for more AI battles. ⚔️

[Latent Operator @LatentOperatorX]

@MarvellousDev Yes the judge scores the video

[Latent Operator @LatentOperatorX]

Same MuseSpark 1.3. DeepSeek Harness scored 8.1/10. Claude Code? Just 4.8/10?! 🤯🏎️

Same rally-in-the-rain prompt. Highest reasoning level (xhigh). Only the harness changes.

DeepSeek Harness: 8.1/10 · 13m51s
Codex: 6.5/10 · 17m34s · ~$1.62
Claude Code: 4.8/10 · 30m49s · ~$0.87

Round 2 of the Harness Cup 🏆 The AI judge picked DeepSeek Harness. Who won for you? Reply 👇

[Latent Operator @LatentOperatorX]

Round 2

[Latent Operator @LatentOperatorX]

Same MuseSpark 1.3. DeepSeek Harness scored 8.1/10. Claude Code? Just 4.8/10?! 🤯🏎️

Same rally-in-the-rain prompt. Highest reasoning level (xhigh). Only the harness changes.

DeepSeek Harness: 8.1/10 · 13m51s
Codex: 6.5/10 · 17m34s · ~$1.62
Claude Code: 4.8/10 · 30m49s · ~$0.87

Round 2 of the Harness Cup 🏆 The AI judge picked DeepSeek Harness. Who won for you? Reply 👇

[Latent Operator @LatentOperatorX]

🏆 THE HARNESS CUP · Round 1/4

Codex, Claude Code or DeepSeek Harness? 🤯🚒

Same MuseSpark 1.3 at its highest reasoning level (xhigh). Same firefighters prompt. Only the harness changes.

Codex: 8.3/10 · 6m01s · ~$0.45
Claude Code: 7.6/10 · 3m20s · ~$0.38
DeepSeek Harness: 7.2/10 · 12m42s · ~$0.81

The AI judge picked Codex. But YOU decide.
👇 Reply with the harness that won. Most votes takes the round.

[Latent Operator @LatentOperatorX]

@NabhaniMehdi Fair point about run-to-run variance! This is a single-run tournament, not a definitive statistical ranking. The goal is to compare quality, speed and cost under identical conditions. Multiple runs would be interesting for a more in-depth benchmark

[Latent Operator @LatentOperatorX]

@fanjiewang @opencode @thdxr Thanks for the update, Frank. Appreciate the transparency 👍

[Latent Operator @LatentOperatorX]

What happened to DeepSeek V4.1 Flash on @opencode? 💀

For the past 3 days, it’s been almost unusable. EXTREMELY slow, and struggling with even basic tasks.

I can’t even run my AI benchmarks anymore.

Am I the only one experiencing this?

@thdxr, what’s going on?

[Latent Operator @LatentOperatorX]

Not sure if this is a Southeast Asia routing issue (I’m currently in Bali), but it’s been like this for the past 2–3 days.

My latest measurements:
• 26–70 tokens/s
• Up to 25s before the first token
• Repeated 429 “server overloaded” errors

Anyone else experiencing this?