Cutting Costs: LLMs Get More Affordable!

The LLM Cost‑Cutter

Weekly briefing on advances that shrink the cost of large language models (Feb 3–Feb 10, 2026)

Pricing announcements from AI labs this week

  • Anthropic launched Claude Opus 4.6 with a 1M‑token context (beta) and kept Opus pricing at $5/M input, $25/M output. No token‑price cut, but more capability at the same rate. (itpro.com)
  • OpenAI is retiring the “chatgpt‑4o‑latest” API model on February 16, 2026. Many teams will migrate to GPT‑5.1, which (per VentureBeat’s comparison) carries lower input pricing than 4o—so some workloads may see cost drops without quality loss. This is a lifecycle/deprecation note, not a new price card. (venturebeat.com)
  • DeepSeek remains the low‑cost floor among major providers: its official API page lists V3.2 at $0.28/M input (cache miss), $0.028/M cached input, and $0.42/M output. (api-docs.deepseek.com)

—

1) Model lifecycle shifts can cut your bill even without official “price cuts”

  • What’s new: OpenAI’s Feb 16 API sunset for “chatgpt‑4o‑latest” nudges developers to newer GPT‑5.1 variants. VentureBeat’s rundown shows GPT‑5.1 input at $1.25/M vs. GPT‑4o’s $2.50/M, with equal $10/M output—so like‑for‑like migrations can lower input costs. Tip: validate quality/latency before switching. (venturebeat.com)
  • Why it compounds: As labs retire older SKUs, you can land on faster, often‑cheaper options (especially for input tokens). Combine with caching and prompt compaction to multiply savings across high‑volume apps.

2) New model, same sticker price: Anthropic Opus 4.6 ups context at no extra cost

  • What’s new: Anthropic’s Opus 4.6 adds a 1M‑token context (beta) and new “Agent Teams,” but keeps Opus pricing at $5/M in and $25/M out. For large docs/code, that’s more work per request without a higher per‑token rate. (itpro.com)
  • Why it compounds: Bigger context reduces chunking/orchestration overhead and failed runs, which lowers total cost per task. If you can do a whole repo or deal in one pass, you spend fewer orchestrator CPU hours and fewer LLM calls overall.

3) Ground truth you can use in negotiations: GPU cloud price cards updated this week

  • CoreWeave (as of ~3 days ago): GB200 NVL72 at $42/hr; B200 8× at $68.80/hr; H200 8× at $50.44/hr; H100 8× at $49.24/hr; RTX Pro 6000 Blackwell 8× at $20/hr. (coreweave.com)
  • Lambda (as of ~2 days ago): On‑demand B200 at $4.99 per GPU‑hour; 1‑Click B200 clusters “as low as” $3.79/GPU‑hr with short commitments. (lambda.ai)
  • Why it compounds: Public, recent reference rates compress margins and help you right‑size fleets (premium GPUs for training; cheaper/consumer‑class for inference). As serving software improves throughput per GPU, your effective $/million tokens drops even if list prices don’t budge.

4) Power price stability: Google inks a 1‑gigawatt solar deal in Texas

  • What’s new: Google signed a 15‑year PPA with TotalEnergies for 1 GW of solar in Texas (Wichita and Mustang Creek projects). Long‑term, fixed‑price energy helps stabilize the power component of AI serving costs. (investors.com)
  • Why it compounds: Steadier, lower‑volatility electricity improves utilization planning, keeps GPUs online during peaks, and reduces the “retry tax” from outages—better dollars per token and higher “intelligence per watt.”

5) Faster paths to firm power inside PJM (policy progress continues)

  • What’s new: FERC accepted PJM’s new fast‑track to swap retiring plants with replacement generation at the same site, and previously ordered PJM to spell out how data‑center co‑location with generation should work. Upcoming tariff filings (due by Feb 16, 2026) will clarify service options for co‑located loads. (insidelines.pjm.com)
  • Why it compounds: Clearer rules for “power next to compute” reduce interconnection delays and wheeling costs, enabling denser racks and higher uptime. Lower soft costs and fewer curtailments mean more tokens per dollar.

6) Facility tech: Microsoft backs high‑temperature superconducting (HTS) power lines

  • What’s new: Microsoft is funding VEIR to apply HTS cables that deliver the same power with about a 10× reduction in cable size and near‑zero line losses. That can shrink electrical rooms and bring more power into the white space. (theverge.com)
  • Why it compounds: Less distribution loss + higher site power density = more GPUs per building and lower facility overhead per token. Pair with liquid cooling and you raise the ceiling on “intelligence per watt.”

7) Fresh research this week that reduces compute and memory per answer

  • Semantic‑aware speculative decoding (SemanticSpec): Verifies the meaning of multi‑token drafts instead of each token, reporting up to 2.7× speedups on some 32B models. Plain English: batch‑check chunks of thought, accept more at once, and move on. (arxiv.org)
  • Hierarchical Adaptive Eviction (HAE) for multimodal KV‑cache: Reduces KV memory ~41% with minimal quality loss and speeds up generation ~1.5× in tests, by evicting the least‑useful vision/text keys differently across the forward and decode phases. Plain English: keep the useful memory, drop the rest—especially vision tokens—so you fit more requests per GPU. (arxiv.org)
  • Why it compounds: KV‑cache traffic is a hidden energy hog. Better eviction and smarter speculation raise tokens/sec on the same cards, cut latency, and let you serve longer contexts without upgrading hardware.

—

Quick takes (this week)

  • China‑led price pressure remains intense; DeepSeek’s official card still shows sub‑$0.30/M input and $0.42/M output for V3.2—use as a negotiation anchor, but verify quality/latency for your workload. (api-docs.deepseek.com)
  • PJM emergency authority during January’s cold snap officially expired February 2; rehearing granted Feb 4. For operators in PJM territory, expect continued regulatory motion aimed at reliability and faster hookups. (energy.gov)

—

Action checklist for the week

  1. If you’re on GPT‑4o in API, dry‑run GPT‑5.1 swap now. Compare input token costs and end‑to‑end latency before Feb 16, 2026. (venturebeat.com)
  2. Re‑benchmark long‑context jobs on Opus 4.6. Bigger windows at the same rate can reduce orchestration overhead (fewer chunked calls). (itpro.com)
  3. Re‑shop GPU capacity with current public cards (CoreWeave, Lambda). Use them to push down committed and spot rates. (coreweave.com)
  4. In PJM, ask about co‑location transmission options and the replacement‑generation fast track; map how these could accelerate your on‑site power. (insidelines.pjm.com)
  5. Trial a semantic‑aware speculative decoder or recent KV‑eviction method on your longest prompts; measure tokens/sec, VRAM, and error deltas. (arxiv.org)

If useful next week, I can add a one‑pager turning these facility, policy, and software changes into an “all‑in $/M tokens” for your current stack.