OpenAI fights back

Summary of Tech/Model Content (GPT 6.1 “Soul” vs others)

1) What models are being discussed and why

• Anthropic released Opus 5.5 recently and the speaker credits it as “unbelievably good,” implying it pushed the market.
• OpenAI allegedly countered quickly with GPT6 Soul and GPT6 Luna, but the speaker says they were not very impressive, so they didn’t cover them.
• The focus is OpenAI’s early-access model “GPT 6.1 Soul,” now available broadly, positioned as:
  • Surprisingly cheap
  • Surprisingly capable
  • Possibly better value than Anthropic’s Astra/Opus depending on task type

2) Benchmarks and cost/performance claims (Terminal Bench, Deep SWE)

The speaker ran Terminal Bench 4 (and mentions Deep SWE) using Terminal Bench on random VMs on their network, with caveats:

• Results vary across runs (bench churn / max runs incomplete).
• Some tasks were removed because they required H100 GPUs.
• The speaker is skeptical of benchmarks and emphasizes broader coverage rather than “perfect benchmarking.”

Terminal Bench 4

• GPT 6.1 Soul scores very highly (often near the top / best among compared models).

Deep SWE

• GPT 6.1 Soul matches Opus while being much cheaper.

Pricing comparison (as claimed)

• GPT 6.1 Soul: ~$21 (low cost tier)
• Astra: ~$1.46 (low tier)
• Opus: ~$14.65 (max tier achieving a similar score, per the speaker’s comparison)

The speaker frames this as “same quality for radically different cash/token cost,” especially considering cash read costs for agent workloads.

3) Pricing details that drive “agentic” affordability

The speaker cites these model pricing details:

• $2 per 1M input tokens
• $10 per 1M output tokens
• Cache read (“cash read”): 10 cents per 1M cached-in, described as a 90–95% discount vs normal reads

They argue the cached-read pricing is the major lever because agent workloads are mostly cache reads (~96%).

They also suggest OpenAI’s subscription math shifted:

• A reference to Tibo’s tweet claims that Pro $200 plan usage calculations were changed, effectively halving API spend vs the prior plan.

Overall implication: OpenAI is cutting margins/subsidies, likely due to competitive pressure from high-performing models like Opus 5.5.

4) Model behavior/quality characteristics (what improved vs GPT 6 Astra)

The speaker contrasts GPT 6 Astra vs GPT 6.1 Soul:

• Astra: “smart and dumb at the same time”
  • Good outputs, but with spiky failures
  • Random spikes into the “dumbest behavior”
• GPT 6.1 Soul: less spiky
  • Still not revolutionary in every domain
  • Slightly below Astra in some advanced “3D/computer use”-type peaks
  • But “good enough” for day-to-day use

Real-world “computer use” trust example

• The speaker describes an email/inbox workflow:
  • Finds unpaid invoices
  • Creates Chrome tabs/workflows for investing wiring
  • Fills details and queues actions
• They claim no mistakes in that demo, making it “trustworthy enough” for tasks like wiring setup.

5) Coding workflow positioning (how the speaker would use it)

Speaker thesis:

• 6.1 Soul excels at review, auditing, and investigation (e.g., “deep code reviews” and root-cause analysis).
• Opus 5.5 remains better as an implementation/collaborator:
  • More pleasant
  • Fewer blocks
  • Better for actually writing/continuing code

Examples

• Deep audits on real repos (e.g., “orchestrator v2”) reported as nearly identical quality to Astra while being cheaper.
• Task: “find opportunities to improve a code base” (T3 improvement mention) scored very well on 6.1 Soul.
• PR review / bug repro / reviewer-style workflows
  • A “10 model bench” blind-style review is mentioned where 6.1 Soul placed first
  • Emphasis on lower runtime cost vs Sonnet/Opus

Critical real-world coding claim

• A TypeScript → Rust port had been stuck for months.
• The speaker claims multiple models were failing/unblocking, but 6.1 Soul boosted progress from ~83.7% to 100% within a day, attributed to:
  • Discovering a large amount of unused code (1.3M of 1.8M lines unused)
  • Explaining how earlier agents’ rewrites left “slop” and unused crates
• They say Opus got the port working but didn’t fully clean up the mess; Soul’s audit uncovered the structural issue.

6) “Efficiency routing” critique (Deep SWE + OpenRouter Jev Router)

The speaker critiques OpenRouter’s Jev Router, described as a cost-optimization strategy using cheaper model dispatch.

Findings:

• The router mostly routes to DeepSeek v4.1 Flash (~60% of requests).
• It ended up more expensive and slower than expected:
  • ~4–5× longer than GPT 6 Astra low tier (example: 4.6 minutes vs 20 minutes)
  • Much higher step count (104 vs 19)

Takeaway: The speaker strongly positions 6.1 Soul as delivering high Deep SWE performance at low cost (example claim: ~4.8 minutes and ~21 cents) outperforming the router strategy.

7) “Fish slop” gaming demo: strengths and regressions

The speaker discusses “Fish slop,” a section about model-driven game generation/simulation.

Strengths

• Visual fidelity looks improved (fish/graphics/sub-details).
• The model is “really good at Blender” tasks:
  • If given CLI/screenshot instructions, it can generate 3D models well.

Regressions

• UI/UX issues:
  • Overly cluttered/garbage UI text
• Gameplay feel:
  • Worse movement (“jank”)
  • Worse frame rate
• They suggest they may not want it generating full game mechanics, despite decent simulation visuals.

8) Final recommendation / speaker’s conclusion

• The speaker says not to replace Opus 5.5 as the daily driver (at least not fully).
• Recommended usage:
  • Use 6.1 Soul by default for scoped work: auditing, deep reviews, and investigation
  • Keep Opus 5.5 as the main implementation/collaboration model
  • Potential hybrid: Opus calls Soul for review/analysis tasks
• They imply a “subsidization era” may be ending due to pricing changes and reduced margins.

---

Main speakers / sources

• Main speaker: The YouTube creator (unnamed in subtitles), speaking about OpenAI/Anthropic models and their own testing.
• External referenced source: Tibo (tweet used to infer subscription/API usage calculation changes).

Referenced products/models

• Anthropic: Opus 5.5, Astra
• OpenAI: GPT 6 Soul, GPT 6.1 Soul
• OpenRouter: Jev Router
• DeepSeek: v4.1 Flash
• Benchmark suites: Terminal Bench 4, Deep SWE