Model routing over model loyalty: splitting workloads across GPT-5.6, Sonnet 5 and Grok 4.5 in 2026

Summary. Three frontier models landed within ten days in mid-2026: Claude Sonnet 5 on 30 June, Grok 4.5 on 8 July, and OpenAI's GPT-5.6 family on 9 July after a limited preview from 26 June. Their list prices per 1M tokens differ by 5x, from $1 input / $6 output for GPT-5.6 Luna to $5 / $30 for GPT-5.6 Sol, with Sonnet 5 at $2 / $10 and Grok 4.5 at $2 / $6. Those numbers are the least reliable input to your decision. Anthropic's own documentation says Sonnet 5's tokenizer turns the same input text into "approximately 30% more tokens than on Claude Sonnet 4.6", which means a per-token price is not a per-request cost. Sonnet 5's $2 / $10 is introductory and rises to $3 / $15 on 1 September 2026. Grok 4.5 has no batch discount at all, while OpenAI and Anthropic both cut 50%. Sam Altman, OpenAI's chief executive, told CNBC on 9 July that GPT-5.6 Sol is "54% more token efficient" on agentic coding, which is a vendor claim and also a concession that tokens, not token prices, are the unit that matters.

What actually shipped, and what it costs

Start with the verified list prices. Every figure here comes from the vendor's own pricing page as of 16 July 2026.

Model	Input per 1M	Output per 1M	Context window	Notes
gpt-5.6-sol	$5.00	$30.00	Not stated on pricing page	Flagship; max reasoning effort and ultra subagent mode
gpt-5.6-terra	$2.50	$15.00	Not stated on pricing page	OpenAI: "competitive performance to GPT-5.5 while being 2x cheaper"
gpt-5.6-luna	$1.00	$6.00	Not stated on pricing page	Cheapest frontier-family tier
claude-sonnet-5	$2.00	$10.00	1M tokens	Introductory; becomes $3.00 / $15.00 on 1 Sep 2026
claude-haiku-4-5	$1.00	$5.00	200k tokens	There is no Haiku 5
grok-4.5	$2.00	$6.00	500k tokens	No batch discount

Run one workload through that table and the spread is obvious. Take 10M input and 1M output tokens a day, a mid-sized production assistant, and hold everything else constant:

Model	Cost per day	Cost per 30-day month
gpt-5.6-sol	$80.00	$2,400
gpt-5.6-terra	$40.00	$1,200
claude-sonnet-5 (to 31 Aug 2026)	$30.00	$900
claude-sonnet-5 (from 1 Sep 2026)	$45.00	$1,350
grok-4.5	$26.00	$780
gpt-5.6-luna	$16.00	$480
claude-haiku-4-5	$15.00	$450

Those are arithmetic from published list prices at a fixed token volume, not measurements of your workload. That distinction is the whole article.

Why the list price is not the price

Four mechanisms sit between the number on the pricing page and the number on your invoice. Three of them are vendor-specific, which is exactly why single-vendor loyalty costs money.

Tokenizers are not a common unit

This is the one almost nobody prices in. A token is not a standard measure. Each vendor tokenises text differently, so "$2 per 1M tokens" at vendor A and "$2 per 1M tokens" at vendor B are prices for different amounts of work.

You do not have to take that on faith, because Anthropic documented it against itself. Sonnet 5 ships a new tokenizer, and the Anthropic documentation states that the same input text produces "approximately 30% more tokens than on Claude Sonnet 4.6". The per-token price did not go up. The per-request cost did.

Now read Sonnet 5's introductory pricing again. Anthropic's launch material says the intro pricing is "set so that the transition to Sonnet 5 is roughly cost-neutral". The $2 / $10 is not a discount. It is compensation for a tokenizer that emits more tokens for the same text. Which means the 1 September 2026 move to $3 / $15 is not a 50% price rise. It is a 50% rate rise on top of a token count that already grew. Budget accordingly.

The same logic cuts the other way for OpenAI's claim. If GPT-5.6 Sol really is "54% more token efficient" on agentic coding as Altman told CNBC, then Sol at $5 / $30 is not straightforwardly 2.5x the price of Sonnet 5 at $2 / $10, because it is billing fewer tokens for the same task. xAI makes a parallel claim for Grok 4.5, saying it has "roughly 2x the token efficiency of comparable leading models, solving tasks in under half the number of steps".

All three are vendor claims measured on vendor-chosen tasks. Treat them as hypotheses to test, not facts to plan with. The point is not that any of them is right. The point is that all three vendors now compete on tokens-per-task rather than price-per-token, which tells you the pricing page has stopped being a comparison tool.

Caching economics differ by vendor

Every vendor discounts cached input. None does it the same way.

Mechanism	GPT-5.6	Claude Sonnet 5	Grok 4.5
Cache read	90% discount ($0.50 on Sol)	$0.20 per 1M intro, rising to $0.30	$0.50 per 1M
Cache write	1.25x uncached input ($6.25 on Sol)	$2.50 per 1M for 5 minutes, $4 for 1 hour	Not stated
Cache lifetime	30-minute minimum, explicit breakpoints	5-minute and 1-hour options	Not stated
Counts toward rate limit?	Not stated	Cache reads do not count toward ITPM	Cached tokens still count toward TPM

OpenAI's June 2026 post is specific: GPT-5.6 "introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30-minute minimum cache life. For GPT-5.6 and later models, cache writes are billed at 1.25x the model's uncached input rate, while cache reads continue to receive the 90% cached-input discount."

The rate-limit row is the one that will surprise your platform team. On Anthropic, cache reads do not consume your input-tokens-per-minute budget, so a heavily cached workload gets effective throughput well above its nominal limit. On xAI, cached tokens still count toward tokens-per-minute even though they are billed less. Two workloads with identical cache hit rates will hit their ceilings at different volumes on different vendors. That is a capacity question, not a pricing question, and it does not appear on any pricing page.

Batch is not universal

Both OpenAI and Anthropic cut 50% for asynchronous batch work. GPT-5.6 Sol drops to $2.50 / $15.00 and Sonnet 5 to $1 / $5 on the introductory rate. Grok 4.5 has no batch discount: xAI's pricing page lists batch discounts for some models and states that models not listed have none, and grok-4.5 is not listed.

If a meaningful share of your volume is offline, nightly enrichment, backfills, evals, classification, then Grok 4.5's attractive $2 / $6 list becomes its only price, while its rivals halve. A workload that is 60% batchable inverts the ranking in the table above.

Priority tiers are not universal either

OpenAI sells priority processing at roughly 2x standard, putting Sol at $10 / $60. xAI sells the same idea at "2x standard rates". Anthropic does not offer a Priority Tier on Sonnet 5 at all.

If your latency commitment requires a priority lane, Sonnet 5 is not a candidate for that workload regardless of its price, and no amount of comparing $2 against $5 will tell you so.

The framework: route by workload shape, not by vendor

Model loyalty is an artefact of procurement, not engineering. One contract, one integration, one bill. It is comfortable, and it means every workload pays flagship rates for work a budget tier would do identically.

Routing means classifying work first and choosing a model second. Four questions do most of it.

Is the task's answer checkable? Classification, extraction, routing, and schema-filling have a right answer you can assert against. These belong on the cheapest tier that passes your evals, which is usually Haiku 4.5 at $1 / $5 or GPT-5.6 Luna at $1 / $6, not on a flagship.

Is it batchable? If it does not block a user, it should be on a batch endpoint at a vendor that discounts batch. That single decision halves the bill for that slice and rules Grok 4.5 out of it.

Does it need the context window, and how much? Sonnet 5 offers 1M tokens and Grok 4.5 offers 500k. For a workload that genuinely needs to hold a large corpus in context, that capability decides the choice before price does. Note that xAI's own model page says it charges different rates above the 200K context window, and the actual rate does not render on the page. Get it in writing before you design around it.

Is it agentic and long-horizon? This is where the token-efficiency claims matter and where you must measure rather than trust. A model that costs 2.5x per token but resolves the task in half the steps is cheaper. It is also the only category where the vendor benchmarks are even arguably relevant.

Workload	Route to	Why
Classification, extraction, tagging	Haiku 4.5 or GPT-5.6 Luna	Checkable output; $450 to $480/month at 10M+1M tokens/day
Offline enrichment, backfills, evals	GPT-5.6 Terra or Sonnet 5, on batch	50% batch discount; Grok has none
Long-context document work	Sonnet 5 (1M) or Grok 4.5 (500k)	Capability gate; price is secondary
Interactive assistant, latency-bound	GPT-5.6 Terra or Grok 4.5	Grok 4.5 is served at "80 TPS"; Sonnet 5 has no priority tier
Long-horizon agentic coding	Measure Sol against Sonnet 5 on your tasks	Vendor efficiency claims conflict; only your traces settle it
Anything user-visible and high-stakes	The tier your evals clear, not the cheapest	Route on measured quality, not list price

What xAI actually published, and what it did not

Grok 4.5's benchmark disclosure deserves a specific note, because it is unusually honest and unusually easy to misread.

xAI published five coding-adjacent results for Grok 4.5: DeepSWE 1.0 at 62.0% pass@1, DeepSWE 1.1 at 53%, SWE Marathon at 29.0% resolution rate pass@1, Terminal Bench 2.1 at 83.3%, and SWE Bench Pro at 64.7%. By xAI's own charts, Grok 4.5 leads on one of those five. The company states that "competitor figures are drawn from the respective developers' published system cards or benchmark leaderboards".

It also publishes no SWE-bench Verified figure for Grok 4.5, which is the benchmark most buyers ask about.

For contrast, we could not verify a single Sonnet 5 benchmark score from Anthropic's own pages: the launch post's benchmark table is published as an image, with the numbers absent from the page text. If a number matters to your decision, note that "the vendor did not publish it in machine-readable form" is itself a finding.

Our own reads on the individual models sit in GPT-5.6 versus Claude Sonnet 5 for enterprise agents, the Grok 4.5 enterprise coding evaluation, and GPT-5.6 inference cost for enterprise AI. To size any of this before committing, our roundup of free tools to measure LLM costs is the cheaper first step.

What routing costs you

Routing is not free, and the honest version of this argument prices the downside.

You now maintain three SDKs, three sets of credentials, three rate-limit regimes, and three failure modes. Your prompts are not portable: a prompt tuned on Sol will not behave identically on Sonnet 5, and the tokenizer differences mean even your context-window budgeting is per-vendor. Your evals have to run per model, which multiplies the judge costs. Sonnet 5 rejects temperature, top_p and top_k outright, returning 400, so a shared request builder needs vendor branches.

The break-even is volume. At $450 a month, routing complexity is not worth anyone's time; run everything on one model and move on. At $2,400 a month for a workload that a $480 tier would serve, the gap pays for a week of platform work in its first month. Do the arithmetic before you build the router.

The real cost of a router is not the code. It is that every model upgrade now has three blast radii instead of one.

India-specific considerations

For teams building from India, two things change the calculation and neither is on a pricing page.

Availability is not uniform. xAI's 8 July announcement said Grok 4.5 was "not yet available in the EU", with availability "expected in mid-July". We found no confirmation either way as of 16 July 2026. Regional availability, and the timing of it, is a design constraint for anyone serving multiple geographies from one router, and it changes without a changelog entry you will notice.

Then there is where the tokens go. A router that sends the cheapest slice of traffic to whichever vendor is cheapest this quarter is also a router that moves customer data across three processors with three sets of terms. Under the Digital Personal Data Protection Act 2023 (DPDP), the processor you send personal data to is a decision you have to be able to defend, and "the router picked it" is not a control. Pin data-bearing workloads to the vendors you have assessed, and let the router optimise only over the traffic where that question is already settled. This is the argument for classifying workloads before you price them.

The one thing to do first

Do not start by choosing a model. Start by instrumenting the workload you already run: tokens in, tokens out, cache hit rate, and what share of volume could tolerate a batch endpoint. Those four numbers turn every table above from a price list into a decision.

Most teams discover two things when they do this. A large share of their flagship traffic is classification work that a $1 tier handles identically. And their cache hit rate is far lower than they assumed, which is where the money actually went.

The vendors have already told you what they are competing on. All three now advertise tokens-per-task rather than price-per-token. Measure the same thing they do.

FAQ

Is GPT-5.6 cheaper than Claude Sonnet 5?

It depends on the tier. GPT-5.6 Luna is $1 input and $6 output per 1M tokens against Sonnet 5's introductory $2 and $10, so Luna is cheaper. GPT-5.6 Sol at $5 and $30 is more expensive. Sonnet 5 rises to $3 and $15 on 1 September 2026, which narrows the gap again.

Why can't I just compare price per million tokens across vendors?

Because a token is not a standard unit. Each vendor tokenises text differently, so the same request bills a different token count at each. Anthropic's documentation states Sonnet 5 produces approximately 30% more tokens than Sonnet 4.6 for identical input text, proving that token counts shift even within one vendor's own lineup.

Does Grok 4.5 offer a batch discount?

No. xAI's pricing page lists batch discounts for certain models and states that models not listed have no batch discount, and grok-4.5 is absent from that table. OpenAI and Anthropic both cut 50% for batch work. If much of your volume is offline, that absence outweighs Grok 4.5's competitive $2 and $6 list price.

What happens to Claude Sonnet 5 pricing on 1 September 2026?

Anthropic's introductory rate of $2 input and $10 output per 1M tokens ends, moving to $3 and $15. Cache read pricing rises from $0.20 to $0.30 per 1M. Because Anthropic says the intro pricing was set to make the Sonnet 5 transition roughly cost-neutral against its new tokenizer, budget for the rate rise and the token count together.

Which model should handle classification and extraction work?

The cheapest tier that clears your evals, normally Claude Haiku 4.5 at $1 and $5 per 1M tokens or GPT-5.6 Luna at $1 and $6. These tasks have checkable answers you can assert against, so quality is measurable rather than a matter of judgement. At 10M input and 1M output daily that is roughly $450 monthly.

Does Claude Sonnet 5 have a priority tier for latency?

No. Anthropic does not offer Priority Tier on Sonnet 5, while OpenAI's priority processing runs at roughly double standard rates and xAI prices Priority Processing at 2x standard. If a workload carries a latency commitment that needs a priority lane, Sonnet 5 cannot serve it regardless of how its per-token price compares.

How large are the context windows on these models?

Claude Sonnet 5 offers 1M tokens, which Anthropic states is both the default and the maximum with no smaller variant. Grok 4.5 offers 500,000 tokens, though xAI notes it charges different rates above 200K context without publishing that rate clearly. OpenAI's pricing page does not state GPT-5.6 context windows.

When is model routing not worth the complexity?

At low volume. Routing adds three SDKs, three credential sets, three rate-limit regimes and per-vendor eval runs, and prompts are not portable between vendors. If your spend is around $450 a month, that overhead costs more than it saves. When a flagship workload runs $2,400 monthly against a $480 alternative, routing pays for itself quickly.

How eCorpIT can help

eCorpIT helps engineering organisations put evidence under their model spend. Our senior engineering teams instrument the four numbers that decide everything above, tokens in, tokens out, cache hit rate and batchable share, then run per-model evals on your actual traces rather than vendor benchmarks, and build the routing layer only where the arithmetic justifies it. We design these systems aligned with DPDP 2023 requirements, pinning data-bearing workloads to assessed processors rather than letting a cost router choose them. If your AI bill is growing faster than your usage, talk to our team.

References

1. OpenAI, Previewing GPT-5.6 Sol: a next-generation model, 26 June 2026
2. OpenAI API pricing
3. Anthropic, pricing documentation
4. Anthropic, What's new in Claude Sonnet 5
5. Anthropic, models overview
6. Anthropic, Claude Sonnet 5 announcement, 30 June 2026
7. Anthropic, API rate limits
8. xAI, Grok 4.5 model documentation
9. xAI, developer pricing
10. xAI, Grok 4.5 announcement, 8 July 2026
11. xAI, developer rate limits
12. CNBC, OpenAI's newest AI model is 54% more token efficient on agentic coding, Altman tells CNBC, 9 July 2026
13. TechCrunch, OpenAI launches its new family of models with GPT-5.6, 9 July 2026
14. Axios, OpenAI releases GPT-5.6 and ChatGPT Work tool, 9 July 2026

Last updated: 16 July 2026.