DeepSeek's Cheap Model Beat Its Own Flagship, Grok Learned to Talk Without Lag, and Claude Sonnet 5 Is About to Cost More Than the Sticker Price Says

Three stories from the last few days aren't about a new frontier model topping a leaderboard — they're about assumptions that quietly stopped being true: which tier in your router is actually the best model, whether full-duplex voice was ever really off the table on latency grounds, and what "cost-neutral" pricing actually costs once you account for how the text gets counted. Here's what happened, and what's worth checking against your own stack.

1. The "Budget" Model Now Beats the Flagship It's Priced Under

DeepSeek moved V4-Flash out of preview on July 31 as DeepSeek-V4-Flash-0731, keeping the same 284B-total/13B-active MoE architecture as the preview build and shipping with only a re-run of post-training — no architecture change. The result: it now scores higher than DeepSeek's own V4-Pro-Preview on all nine agent and coding benchmarks the company has published, including a jump on Terminal-Bench 2.1 from 61.8 to 82.7. Input pricing held at $0.14 per million tokens, the same rate as before the retrain.

The detail worth sitting with isn't that a "flash" tier model closed the gap with its flagship sibling — it's that nothing about the model's category changed to make that true. A post-training pass moved the ranking a router built weeks ago no longer reflects, and nothing in a model name or price tier tells you that happened.

const modelRouter = {
  simpleTasks: 'v4-flash',   // priced and ranked as the cheap tier
  hardTasks: 'v4-pro',       // assumed to be the better model
  lastRerankedAgainstBenchmarks: null, // the assumption nobody re-checked
};
// "Pro" in a model name is a label from launch day,
// not a live guarantee that it's still the stronger model.

If your routing logic sends hard tasks to a "pro" or "flagship" tier by name rather than by current benchmark standing, this is a good week to re-run that comparison — a cheaper tier may already be outperforming the one you're paying more for.

2. Full-Duplex Voice Crossed a Line It Was Stuck Behind for a Year

xAI shipped Grok Voice Think Fast 2.0 on July 31, and the headline number isn't the top-line quality score — it's the Full Duplex Bench subset, which jumped from 77.8% on Think Fast 1.0 to 95.1% on 2.0. Overall, the High variant lands at #2 on the Artificial Analysis Speech-to-Speech Index at 82.9% (behind Qwen Audio 3.0 Realtime Plus at 84.1%, ahead of GPT-Realtime-2.1 High at 79.1%), and #1 on Tau Voice's agentic-performance benchmark at 56.5%. Time-to-first-audio averages 0.70 seconds — the only model in the index's top five under one second. Pricing lands at $4.80 per hour of input audio, roughly 2.2x cheaper than GPT-Realtime-2.1 High's $10.75.

The number that matters for builders isn't the ranking, it's the 17-point jump on the one subset that measures whether a model can be interrupted mid-sentence and correctly hand the turn back without stepping on itself. That's the specific failure mode that has made most teams default voice products to strict push-to-talk instead of open-mic conversation — and it just moved further in one release than it had over the prior several.

const voiceAgentConfig = {
  turnTaking: 'push-to-talk', // chosen because duplex used to feel laggy/broken
  fullDuplexBenchAtLastReview: 77.8,  // Think Fast 1.0
  fullDuplexBenchNow: 95.1,           // Think Fast 2.0, July 31
};
// A UX decision made against last year's duplex score
// is worth re-testing against this year's, not assumed permanent.

If open-mic, interruptible voice was ruled out for your product because duplex handling felt unreliable, it's worth a fresh test — the specific number that drove that call has moved substantially.

3. "Cost-Neutral" Pricing Isn't Neutral Once You Count the Tokens

Claude Sonnet 5's introductory pricing — $2 per million input tokens, $10 per million output — holds through August 31 and moves to standard pricing of $3 and $15 on September 1, a 50% increase on the sticker rate. Anthropic frames the new standard rate as cost-neutral, since it matches Sonnet 4.6's existing rate card. But Sonnet 5 runs on a new tokenizer that counts roughly 30% more tokens for the same piece of text than 4.6's tokenizer did — so matching the per-token price does not mean matching the per-request bill.

The gap worth checking isn't the well-publicized 50% rate change, it's the unpublicized multiplier underneath it. A cost projection built by taking a known prompt, running it through Sonnet 5 today, and multiplying by 1.5 for September will still be wrong — it needs to account for the extra ~30% in token count on top of the rate change, not instead of it.

// Naive projection: only catches the sticker-price change
const naiveCost = tokensAtSonnet5Rate * 1.5; // wrong

// What actually changes on September 1
const actualCost = (tokens * 1.30) * 1.5; // tokenizer inflation × rate change

If your budget model for Sonnet 5 was built on intro pricing and a flat percentage bump, re-run it against actual token counts from the new tokenizer before September 1 — the compounding, not the headline number, is where the surprise is.

What This Means for Builders

None of this week's stories is about a benchmark record. The DeepSeek release says a model's tier label is a snapshot from launch day, not a live ranking — worth re-verifying before you keep routing hard tasks to the "flagship" by name. The Grok release says a UX call made against last year's duplex score deserves a retest now that the number behind it moved 17 points. And the Sonnet 5 pricing change says a rate-card comparison can look cost-neutral while a tokenizer change quietly compounds on top of it. Check your own stack against whichever of the three actually touches what you're shipping.

— Maya