Oka Si

Which LLMs I use as of August 2026

My current model and provider mix after Fable and GPT-5.6: rejections, Hermes for memory, my own proxy for routing, subscription plans with ZDR, and the RTX 3090 plan.

Published

Last edited

The frontier moved again. Fable and GPT-5.6 shipped, and with them came a familiar pattern: more requests start getting rejected & flagged. Rate limits, capacity gates, silent denials, whatever the mechanism, it is frustrating enough that I have been rebalancing the stack.

As of August 2026, my stack is reliability-first: Grok 4.5 is the frontier intelligence pick when I want more peak reasoning, not the always-on path.

Hermes + my own proxy

I mostly run through Hermes Agent for persistent memory and long-lived threads. Hermes is the agent runtime: sessions, memory, tools.

Routing sits in my own proxy. Hermes talks to that proxy; the proxy fans out to providers and models depending on the job: coding agents, research, bulk extraction, or chat that needs to survive across days.

That split matters. Memory and continuity stay with Hermes; capability, price, and failover get decided in the proxy without throwing away context.

What I actually run

RoleModelWhy
Frontier intelligenceGrok 4.5Peak reasoning when I want the smarter pass
High-volume / cheap bulkDeepSeek-V4-Flash-0731Credits stretch farthest; huge request volume
Strong coding volumeGLM-5.2Solid multiplier on subscription tiers
Backup / failoverGPT-5.6 LunaQuality when capacity allows, not the always-on path

Subscription plans & models (where I buy tokens)

I stick to subscription plans rather than pure pay-per-token. Subscriptions are usually subsidized: the list-price value of the tokens you get often exceeds what you pay, which keeps my costs lower than paying per million tokens.

Where to start

  1. OpenCode Go: $10/mo ($5 first month).1 The first plan to get: cheapest entry, generous limits, and one of the three providers with ZDR (Zero Data Retention).2
  2. Command Code GOAT: $10/mo with $70 in credits → ~7× value.3 Best low-cost subscription in the market right now.4
  3. Cursor Pro+: $60/mo.5 For daily agent users; the strongest Grok access through Cursor's Models pool.6

These three are the providers that support ZDR (Zero Data Retention) and still come with value multiples or bonuses, so they're the plans I'd take before anything else.264

ModelCommand Code Go ($1)Command Code GOAT ($10)OpenCode Go ($10)Cursor Pro+ ($60)
GLM-5.2Value $10 · 10×Value ~$70 · ~7× · ~75K req/mo3,200 req/5hLimited
GPT-5.6 LunaValue $10+ · 10×+ (heavy discounts)Value ~$70 · ~7×4,100 req/5hValue ~$70 · ~1.2×
DeepSeek-V4-Flash-0731Value $10 · 10× · credits stretch farthestValue ~$60 · ~6×63,300 req/5hLimited
Grok 4.5Value $10 · 10×Value high · 7×120 req/5hValue very high · high multiplier (Cursor Models pool) · 100M+ tokens possible

Quick pairing suggestions

  • Command Code Go → DeepSeek-V4-Flash-0731
  • Command Code GOAT → GLM-5.2
  • OpenCode Go → GPT-5.6 Luna
  • Cursor Pro+ → Grok 4.5

Next week: local workhorse on an RTX 3090

Cloud is flexible until it isn't. Starting next week I plan to stand up the RTX 3090 for local inference of Qwen 3.8 27B. The bet is simple: a 24/7 workhorse that is financially feasible and reliable enough that rejections stop being a daily tax.

If that holds, Hermes keeps the memory and threads; the proxy can point heavy jobs at the 3090 as the always-on backend for long-running, high-volume, "just keep working" work. Cloud models stay in the mix for peak quality and for tasks that still need a frontier edge.

Practical takeaway

  1. Route, don't marry a single API. Hermes for memory and threads; my proxy for multi-provider routing so threads stay alive when one vendor flakes.
  2. Optimize for rejects as much as for quality. Pick reliability first; Grok 4.5 is for frontier intelligence, not volume.
  3. Buy tokens where the multiplier matches the model. GLM-5.2 on Command Code when you need it; Grok on Cursor's model pool; GPT-5.6 on OpenCode Go when you need it.
  4. Local is the reliability floor. Qwen 3.8 27B on a 3090 is the next step toward a stack that does not depend on someone else's capacity queue.

I'll update this as the 3090 path lands.

Footnotes

  1. OpenCode Go pricing: $5 first month, then $10/month. See opencode.ai/docs/go and opencode.ai/go. ↩

  2. OpenCode Go privacy: model training not used, data retention 30 days (0 days for several models). See the "Privacy" section of opencode.ai/docs/go. ↩ ↩2

  3. Command Code GOAT: $10/month + processing fee, $70 in credits included, up to ~$100 usage with deals, ~75K requests. See commandcode.ai/pricing. ↩

  4. Command Code zero data retention option. See commandcode.ai/privacy. ↩ ↩2

  5. Cursor Pro+: $60/month (USD), recommended for daily agent users. See cursor.com/pricing. ↩

  6. Cursor Privacy Mode: we will not train on your data; technical controls + contractual requirements with model providers. See cursor.com/security. ↩ ↩2

Static sites that still feel alive

Prerendered HTML plus focused client-side motion, image placeholders, and why SSG is enough for a personal site.

On this page

Hermes + my own proxyWhat I actually runSubscription plans & models (where I buy tokens)Where to startQuick pairing suggestionsNext week: local workhorse on an RTX 3090Practical takeawayFootnotes