flocode · a thought piece on token economics

A Token Is Not a Token

The industry quotes “price per million tokens” as if a token were a bushel of corn. It isn’t. A token is a differentiated good along four dimensions — quality, speed, cost, and a fourth that nobody prices: whether the model behind it will still be available to you, on hardware you can get, at a price you didn’t have to accept.

The flocode team · July 6, 2026 · all prices as of 2026‑07 — per Epoch AI, constant-capability token prices fall 9–900× per year; date-stamp everything.

1 · The $0.11 token and the $50 token

Here are two prices, both real, both current. Anthropic will sell you a million output tokens from Claude Fable 5, the best coding model in the world, for $50.1 Anthropic pricing page, 2026-07: Fable 5 $10 in / $50 out, cache hits $1. OpenAI: GPT-5.5 $5/$30, GPT-5.5-pro $30/$180. An 8-GPU AMD MI300X node that we own outright — $160k of capex, depreciated over three years, plus power and colo — costs $8.37 an hour to run, and at saturated batch throughput measured on identical silicon it produces a million tokens of frontier-class open-weight output for about eleven cents.2 $8.37/hr ÷ (21,000 tok/s × 3600 s) = $0.11/Mtok. The 21k tok/s aggregate figure is a vendor-published measurement (Moreh, DeepSeek-class MoE on 8×MI300X); our own conservatively measured c16 figure is ~1,000 tok/s → $2.33/Mtok.

That gap — 455× at the extremes, 21× on our own conservative measurements — is not a scandal. The models differ in quality; batching differs; the API is doing real work you would otherwise do yourself. Some of the spread is honest.

What is strange is that the market prices none of the structure inside the spread. Every provider quotes one number, dollars per million tokens, as if tokens were fungible. They are not fungible. They differ in the quality of the model that produced them, the speed at which they arrive, the cost it took to make them — and in a fourth property the invoice never mentions: whether the model can be re-hosted on other silicon — AMD, a TPU, a wafer-scale engine, the Mac on your desk — tomorrow, next year, or the day after the vendor deprecates it. Call it hardware optionality. The first three dimensions have prices, however crude. The fourth trades at zero.

This essay argues that’s a mispricing, shows who is already quietly paying billions to correct it, and proposes a way to write it down: a Token Grade, G = (Q, L, P, H), with the fourth coordinate scored like the bond rating it deserves to be.

2 · Three dimensions the market already found

Quality was priced first — the mini/nano ladders, the flagship tiers. But the unit of account is already fraying. The current Claude generation’s tokenizer emits roughly 30% more tokens for the same text than pre-4.7 Claude models — Anthropic says so on its own pricing page — which quietly inflates every cross-generation and cross-vendor comparison.3 Anthropic: “Claude Opus 4.7 and later… use a newer tokenizer… approximately 30% more tokens for the same text.” Note this does not stack on the Fable-vs-Opus-4.8 comparison — both use the new tokenizer. Meanwhile hosts serve the same open-weight model at fp8 or fp4 without a standardized disclosure: quantization is a hidden quality tier, sold at the same shelf position.

Speed went from unpriced to three-tiered in about eighteen months. Every major provider now sells the identical −50% batch discount. OpenAI sells priority processing at ~2.5× list. Anthropic sells Fast Mode at 2× — down from 6× one generation earlier. And this month DeepSeek introduced the first true spot signal in the token market: peak-hour surge pricing, output prices doubling during Beijing business hours, framed explicitly as GPU demand management.4 OpenAI priority tier $12.50/$75 for GPT-5.5; Anthropic Fast Mode Opus 4.8 $10/$50 (Opus 4.7 fast was $30/$150, retired 2026-07-24); DeepSeek surge effective mid-July 2026, per TNW.

Cost is the dimension collapsing under everyone’s feet. Epoch AI measures the price of constant capability falling between 9× and 900× per year depending on the benchmark, with a median around 50× and an accelerating ~200× since early 2024.5 Epoch AI, “LLM inference price trends.” GPT-4-level GPQA performance specifically: ~40× cheaper per year. GPT-4 launched in March 2023 at $60 per million output tokens; GPT-4-class output today costs a quarter. But look at the top of the market: GPT-5.5 lists above GPT-5.4. Fable 5 lists at 2× Opus 4.8. GPT-5.5-pro, at $180, costs more than 2023 GPT-4. Yesterday’s intelligence deflates like DRAM; today’s best holds firm. That is not a commodity market. That is a market with grades.

3 · The dimension nobody prices

Now the fourth coordinate. A closed-model token is produced on exactly one fleet, owned by exactly one vendor, on hardware you cannot see, under an allocation you cannot audit, with a deprecation schedule you do not control. An open-weight token can be produced by anyone with compatible silicon — and in 2026 “compatible silicon” is no longer a euphemism for NVIDIA.

The clearest evidence is a market that already exists. GLM-5.2, a frontier-class open-weight MoE, is sold today by seven independent hosts at prices from $0.93 to $3.00 per million input tokens — a 3.2× spread, on three different quantizations, with a visible competitive floor.6 Snapshot 2026-07-03 (digitalapplied): Z.ai and Fireworks $1.40 in; DeepInfra $0.93 (fp4, 32k output cap); GMI $0.98 (fp8); Novita $1.09; Wafer $1.20; Wafer Fast $3.00. One aggregator’s snapshot — spot-check before quoting. That market structure — multiple sellers, one product, price discovery — is impossible for a closed model. There, one seller writes the price sheet, and the “market price” is whatever it says this quarter.

A closed-model token embeds an unpriced short position on one vendor’s fleet, allocation, power queue, deprecation schedule, and price sheet.

So define the grade properly. Q, quality: the capability class, with a precision disclosure (fp8/fp4), because quantization is a quality tier. L, latency: serial tokens per second and time-to-first-token, in tiers the market already uses. P, price: dollars per million, dated, because it decays. And H, hardware optionality, which we score 0–100 as H = V × M × E × 10:

• V — vendor breadth (0–10): how many independent, customer-accessible silicon supply chains can serve this model at production quality? Closed frontier: 0 — you buy tokens, not the right to make them. Open frontier MoE: 6–8 (NVIDIA, AMD Instinct, TPU via vLLM, Trainium, Groq, Cerebras, Ascend, Apple Silicon).
• M — software maturity (0–1): 1.0 means first-party recipes and day-0 support on ≥2 vendors. The measured decay constant of the CUDA moat: DeepSeek’s 1.6T-parameter V4 ran at 1–2 tok/s on AMD’s MI355X on day 0 and beat the H200 by day ~50.7 SemiAnalysis, “DeepSeek V4: day 0 to day 43.” Price new architectures on AMD with a ~1-month maturation-lag discount, not a lock-in discount. ROCm now passes 93% of vLLM’s CI; GLM, Qwen, and DeepSeek ship first-party AMD recipes; EAGLE/MTP speculative decoding runs on MI300X.
• E — edge viability (0–1): 1.0 means it runs on a consumer box (gpt-oss-120B does 43 tok/s on a $2,348 Strix Halo mini-PC — which ties NVIDIA’s $4,699 DGX Spark, because decode is memory-bandwidth-bound and the moat doesn’t reach the edge). 0.5 means Mac-Studio-class. 0 means no weights.8 Strix Halo / DGX Spark: Tom’s Hardware & The Register benchmarks. DeepSeek-class 671B MoE reportedly runs 4-bit at >20 tok/s on a $9,499 512GB M3 Ultra Mac Studio.

Letter grades for contracts: H4, open weights, multi-vendor proven. H3, open weights, NVIDIA-only in practice. H2, closed, but the vendor demonstrably runs multi-substrate — the option exists, and it belongs to the vendor, not to you. H1, closed, single fleet. Every closed frontier model scores a buyer-side H of zero by construction; the letter only records whose balance sheet the option sits on.

Two grade cards make the purchase decision concrete — the same way a nutrition label made calories concrete:

4 · What the smartest money already pays

If hardware optionality were worthless, the people with the best information would not pay for it. Here is what they actually paid, in the last nine months:

• OpenAI took ~10% of AMD as the price of a second source. The October 2025 deal: 6 GW of Instinct GPUs starting with the MI450, and AMD issued OpenAI a warrant for 160 million shares — roughly a tenth of the company — vesting as the gigawatts deploy.9 AMD IR, 2025-10-06; warrant mechanics ($0.01 exercise, share-price milestones) in the SEC 8-K exhibit. NVIDIA countered within weeks: up to $100B invested progressively against a 10 GW NVIDIA-systems deal.
• Then Meta did the same thing. February 2026: up to 6 GW of custom MI450-based inference silicon, and a second performance warrant for 160 million AMD shares. Two of the largest compute buyers on Earth, each paying with ~10% equity stakes, for the identical property: the ability to not depend on one supplier.10 AMD press release 2026-02-24 and SEC 8-K. Anthropic, meanwhile, triple-sources: ~1M Google TPUs (>1 GW in 2026, ~3.5 GW next-gen from 2027) plus up to 5 GW of Amazon Trainium plus NVIDIA.

Hold the two facts side by side. The vendors of closed models pay tens of billions of dollars — in equity, no less — to secure hardware substitutability for themselves. Their API customers receive none of it. The option is real, it is expensive, and it is being bought at scale; it just never appears on your invoice, because on your invoice its price is zero and its quantity is zero.

And the scarcity that makes the option valuable is structural, not cyclical: TSMC’s CoWoS advanced packaging is effectively sold out into 2027 with NVIDIA holding the majority of it; HBM is fully allocated and now 30–40% of an accelerator’s bill of materials; Microsoft has disclosed an $80B backlog it cannot serve for want of power; PJM, the largest U.S. electricity market, just cleared record capacity prices two years running, driven by data-center load.11 PJM 2026/27 BRA: $329.17/MW-day (record); 2027/28: $333.44, the FERC cap, 6,623 MW short of the reliability requirement. PJM attributes ~5,500 MW of peak-load growth “mainly to data centers.”

5 · Pipeline gas, sour crude, and spinning reserve

Every mature commodity market eventually priced its version of this. The analogies are not decoration; they are the operating manual.

From oil, take grades. WTI, Brent, and Dubai are “one commodity” with permanent, exchange-priced differentials for quality and logistics. The lesson is that differentials are prices that move: sweet–sour spreads compress when refineries add desulfurization capacity. Agent scaffolding, verifiers, and test harnesses are the desulfurization units of the token market — they let a buyer run cheaper, sourer crude and still ship clean product. Watch the frontier premium compress accordingly.

From electricity, take the market structure. Tokens, like power, are non-storable, quality-tiered, and locationally priced — the industry already sells geography multipliers, +10% for US-only inference. Prompt caching is the pumped hydro (DeepSeek sells cache hits at ~2% of the miss price). Batch APIs are baseload; Groq and Cerebras are spinning reserve. And above all: electricity runs capacity markets, which pay generators today for the promise of being available in 2027 — the purest possible market price on future availability. PJM derates any single generator with correlated outage risk. A closed model is exactly that: one unit, one owner, whose price hikes, deprecations, throttles, and policy changes are all correlated forced outages. An open-weight model is a diversified portfolio across six or more independent supply chains. Token buyers should derate H1 capacity the way PJM derates a lone gas peaker. Nobody does — yet.12 The recursion is exquisite: AI datacenter demand is already priced as a future-availability product in electricity, at record levels, while AI’s own output has no equivalent market.

From LNG, take the direct precedent. Pipeline gas and LNG are the same molecule. Destination-flexible LNG — the right to land the cargo wherever netbacks are best — grew from ~15% to roughly half of contracted volumes, and the econometrics confirm the flexibility carries positive, measurable option value. Closed tokens are pipeline gas: one route, one counterparty, relationship pricing. Open-weight tokens are destination-flexible cargoes. OpenRouter is the Henry Hub forming in real time — and Chinese open-weight models are already more than 45% of its token volume, up from under 2% a year ago.13 OpenRouter blog, “The open-weight models that matter,” June 2026.

6 · February 1982

One historical episode should hang over every AI procurement meeting. In February 1982, IBM refused to design Intel’s 8088 into the PC unless Intel licensed a second manufacturing source. The era’s biggest buyer would not sole-source a critical input, period. Intel complied. The second source it was forced to license was a small company called Advanced Micro Devices.

AMD exists as a company of consequence because the 1982 buyer refused single-supplier risk — and in 2026, AMD is again the second source, this time as the substrate that makes open-weight tokens supplier-independent, and the two biggest buyers of compute just paid ~10% of the company each to secure it. The 1982 lesson generalizes cleanly: the second source rarely needs to win; it needs to exist, and its existence is what disciplines the incumbent’s price. GLM-5.2’s seven hosts discipline each other to a $3.00 floor. Fable 5’s price is disciplined by nothing but Anthropic’s judgment.

For a closed frontier model, no second source can exist even in principle. The correct 1982 analogy for a closed model is a CPU that IBM would have refused to design in.

7 · “But nobody actually switches”

The objections deserve their strongest form.

“Quality dominates; nobody leaves the best model to save money.” Largely true at the frontier — which is exactly why frontier prices rise into a 200×-per-year deflation everywhere else. But price the quality gap honestly: Epoch’s tracked lag between open weights and closed SOTA is now about three months, the smallest ever; DeepSeek V4-Pro sits within 0.2 SWE-bench points of Opus 4.6, under an MIT license. And the risks accepted for that sliver are no longer hypothetical: Anthropic retired eight Claude models in twelve months; OpenAI’s largest-ever deprecation waves land this month and in October, and fine-tuned models die with their base models; Claude Code was rate-limited at peak hours during an acknowledged compute shortage.14 OpenAI deprecations page (waves 2026-07-23 and 2026-10-23; Assistants API removed 2026-08-26). Anthropic model-deprecation table and formal “deprecation commitments” (≥60 days notice). Rate limiting is deprecation in a different costume. You are not paying 10–100× for three months of quality. You are paying it for quality plus a short option you never see on the invoice.

“The optionality is theoretical — CUDA is a moat.” It was true, and it is now false for inference, with the residue nameable: FlashAttention-3, native NVFP4 checkpoints, DeepEP tuning, NVL72 rack-scale. Everything else — MLA attention, fused MoE, fp8, speculative decoding — has crossed over. ROCm passes 93% of vLLM’s CI; AMD ships day-0 recipes for Qwen and Gemma releases; the correct discount for a brand-new architecture on AMD is a ~1-month maturation lag, measured, not a lock-in discount. Honesty requires the friction, too: numerical behavior diverges across accelerators, and switching cost is real — falling, but real. It belongs in the option price, not in place of it.15 On divergence: arXiv:2511.11601. On the day-0-to-day-50 crossover: SemiAnalysis DeepSeek-V4 tracking. Fairness note: SemiAnalysis also found the moat alive for training in 2024–25 — this essay’s claim is about inference.

“Self-hosting is a fantasy; utilization kills you.” The strongest objection, and partly correct — so let us put our own numbers against it. On our owned MI300X node, a single interactive stream costs about $31 per million tokens: more than GPT-5.5. Below roughly 35% utilization, owned hardware loses money on depreciation alone. Self-hosting one developer’s chat is not the trade. The floor only opens with concurrency — $2.33/Mtok at a measured c16, ~$0.11 at saturated batch — which is to say: the API providers’ entire cost advantage is that they batch you with strangers. But the rebuttal to the objection is the dual-sourcing literature’s oldest result: the option does not require exercise to pay. The second source disciplines price by existing. Seven hosts hold GLM-5.2 at $3; everyone’s standing ability to self-host is what holds the seven hosts. You hold the H4 token not because you will rack GPUs tomorrow, but because the fact that you could is priced into what you pay today — and for whoever does own the metal at fleet utilization, the floor is real.16 Cost floor formula: $/Mtok = ($/hr × 10⁶) ÷ (tok/s × 3600). Owned node: $8.37/hr. At 75 tok/s (c1): $31.00. At ~1,000 tok/s (c16, measured): $2.33. At 21,000 tok/s (saturated batch, vendor-measured on identical silicon): $0.11. An earlier internal draft misplaced this decimal by two orders of magnitude; the numbers here are the corrected ones. Trust, but recompute.

8 · The market that prices it

The infrastructure for pricing H is being poured this quarter. CME Group and Silicon Data announced the first GPU compute futures in May; ICE and Ornn followed a week later — both pending regulatory approval, both cash-settled against GPU rental-rate indices. A compute future hedges the input cost of an open-weight token: rig, rate, throughput, floor. It cannot hedge a closed-model token at all, because a closed model’s price is a discretionary vendor decision uncorrelated with any tradable index. Hedgeable versus unhedgeable is about to become a visible token attribute.

Try the arithmetic yourself — this is the entire cost floor, three inputs:

Four predictions, falsifiable within 24 months. One: compute futures trade, and open-weight operators start writing hedged fixed-price token forwards — closed vendors cannot match without exposing internal costs. Two: a token capacity market emerges: guaranteed Mtok/s-month, auction-cleared, first over open weights. Three: enterprise RFPs begin requiring an H-rating or second-source clause for AI suppliers — the 1982 move, replayed in procurement boilerplate. Four: the frontier premium compresses toward verified-quality-gap × task-value as scaffolding improves, exactly as sweet–sour crude spreads compressed when refineries upgraded.

9 · Building for a multi-grade world

What does a rational actor do in a four-dimensional token market? Stop treating model choice as a config value set once and forgotten. Treat it as portfolio allocation across grades: frontier API tokens where the task genuinely earns the quality premium; open weights on owned or rented substrate where the floor lives; local models as the tranche no counterparty can deprecate, throttle, or reprice.

That is the design premise of flocode, a terminal AI coding agent built for exactly this world — it routes work across frontier APIs, self-hosted open weights on AMD metal, and fully local models, per task, as a first-class decision. We built the tool we’d want to be holding when the market starts stamping H-ratings on tokens. That’s the whole pitch; the essay is the argument.

As far as we can find, no one has yet proposed pricing hardware optionality as a dimension of the token itself — quality, speed, and cost each have a formal literature, and adjacent work exists on tokens as differentiated goods and on lock-in, but the fourth axis is open white space.17 Nearest prior art: “Tiered Super-Moore” (arXiv:2603.28576, tokens as vertically differentiated goods); Grogan (arXiv:2604.06217, provider lock-in); Xing (arXiv:2603.21690, token futures assuming away hardware heterogeneity); SemiAnalysis on the renter/owner cost wedge. A negative existence claim can never be fully confirmed — hence “as far as we can find.” It will not stay open. Markets abhor an unpriced risk the way nature abhors a vacuum — slowly, then all at once.

A GGUF file on your own disk is a token supply contract with no counterparty.

A token is not a token. Grade yours.

© 2026 flocode · written to be printed · argue with the grades: that is what they are for