There is a particular kind of sentence that appears in Chinese and international coverage of cell therapy, and it goes roughly like this: the patient received a fourth-generation CAR-T, and the tumour disappeared. The second half of that sentence is a fact. The first half is a design label, and the two have almost nothing to do with each other. A generation number tells you how many signalling modules were welded into the receptor. It does not tell you how well the resulting cell will work in a given disease, and on at least one occasion in this field, adding a module made the cell worse.
The reason is visible in the historical record. A review published in Nature Reviews Immunology in July 2026 walks the whole arc, from the experiments that identified the T cell in the early 1960s to the in vivo engineering platforms now entering trials. It is written as a story of progress, which is what reviews usually are. But read the same sequence as a list of things that went wrong and were repaired, and it becomes more useful — because it explains why each advance arrived when it did, and what each one cost.
The pattern is consistent enough to state in one line. Every generation of CAR solved the previous generation's failure mode and introduced a new one. Specificity arrived before the cells could survive. Survival arrived with a choice between two kinds of toxicity. Adding more signal produced more exhaustion. Adding cytokines produced control over the microenvironment at the price of systemic inflammation. Adding logic gates produced precision, and precision does not by itself produce a remission.
The short version
The genealogy: five generations of receptor design in roughly twenty years of clinical work, built on four decades of basic immunology that had no therapeutic intent at the time.
The pattern: each generation fixed the previous one's dominant failure and created a new one. More signalling modules is not more efficacy; the third generation, with two co-stimulatory domains, largely failed to beat the second in the clinic.
The most useful historical fact for a patient: the oldest long-term durability data in the field does not come from cancer at all. It comes from HIV trials begun in the 1990s, where engineered T cells were still detectable more than a decade later.
What it means: ask which generation was used only as background. The questions that decide an outcome are the target, the tumour burden, the line of therapy and the safety plan.
Why "Fourth-Generation CAR" Is Not a Quality Rating
The numbering convention is a convenience invented by reviewers, not a regulatory category. No regulator approves a therapy on the basis of its generation, and no trial is designed to compare generation four against generation five. What the numbers actually track is the intracellular signalling architecture: how many co-stimulatory domains, whether a cytokine payload is built in, whether the receptor can be switched on and off.
That architecture matters enormously for how a cell behaves. It matters much less for the question a patient is asking. Two products with identical generation numbers can differ in antigen, affinity, manufacturing process and dose, and those four variables do more to determine an outcome than the signalling cassette does. When a news story leads with the generation, it is leading with the one variable that is hardest to interpret.
The third generation is the clearest evidence for this. It was designed on an entirely reasonable premise — if one co-stimulatory domain helps, two should help more — and it did not work out that way. Constructs carrying both CD28 and 4-1BB eradicated large established tumours in mouse mesothelioma models. In patients, the generation lagged behind the second, with inconsistent preclinical-to-clinical translation, contested benefit in blood cancers, and added concern about cytokine toxicity and premature T cell exhaustion. The design was more powerful on paper and less useful in practice. That is the fact that should attach itself to every future claim about a higher generation.
Before Anyone Imagined a Therapy
The foundations were laid by people asking questions about how the immune system distinguishes self from non-self, with no clinical product in mind. Jacques Miller's neonatal thymectomy experiments in the early 1960s established that T cells were a distinct lineage derived from the thymus. Rolf Zinkernagel and Peter Doherty then showed, using a lymphocytic choriomeningitis virus model, that cytotoxic T cells could only kill infected target cells sharing the same MHC haplotype — the finding called MHC restriction, which established that T cells read an antigen-MHC complex as a single unit rather than reading antigen alone.
By 1982 James Allison and colleagues had identified a disulphide-linked heterodimeric glycoprotein on mouse T cells as the presumed antigen-specific receptor; by 1984 Tak Mak's and Mark Davis's groups had independently cloned the cDNAs encoding T cell-specific subunits, providing the first molecular evidence for the T cell receptor. The TCR-CD3 complex was defined as an octameric signalling unit. In parallel, CD28 was identified as the principal co-stimulatory receptor, and the requirement for a second signal alongside antigen recognition became the organising principle of T cell activation.
Two consequences followed directly. The first was that a receptor could be rebuilt: if specificity lived in a cloned molecule, it could be transferred. The second was that a receptor alone would never be enough, because activation requires co-stimulation — which is precisely the problem that the second generation existed to solve, fifteen years before anyone solved it.
Generation One: Specificity Without Stamina
The first chimeric receptors fused an antibody-derived single-chain variable fragment to the intracellular signalling domain of CD3ζ. The architecture is still the backbone of every approved product: an scFv that binds antigen without MHC presentation, a hinge, a transmembrane region, and a signalling tail. The Eshhar group's constructs, initially called T-bodies, demonstrated that a single polypeptide could redirect a T cell to a chosen target in an MHC-independent way.
They also demonstrated the limit. With CD3ζ alone as the signalling domain, activation was suboptimal and antitumour efficacy was limited. Cells could find their target and could not sustain the fight. In the first cancer trials, reported in 2006, this showed up as two distinct problems: a first-generation carbonic anhydrase IX construct in renal cell carcinoma produced on-target, off-tumour liver toxicity, and a first-generation folate receptor construct in ovarian cancer produced poor persistence. The field learned both lessons at once — that specificity is not the same as safety, and that a cell which cannot persist cannot help.
Generation Two: The Split That Still Divides the Market
The second generation added a co-stimulatory domain, and the choice of which one turned out to shape clinical behaviour for the next two decades. CD28 and 4-1BB are not interchangeable parts. They produce different cells.
CD28-based receptors signal through PI3K-AKT, NF-κB, NFAT and AP-1, and produce rapid activation, vigorous target lysis and heavy cytokine release. 4-1BB-based receptors signal through TRAF-mediated MAPK and NF-κB cascades, and produce a more gradual activation programme that favours persistence, oxidative metabolism and central memory differentiation. In practical terms, CD28 constructs tend to expand faster and produce high-grade cytokine release syndrome earlier; 4-1BB constructs tend to persist longer and produce toxicity on a slower curve.
This is the single most useful piece of design information a patient can hold, because it predicts the shape of the first two weeks rather than the size of the response. It is also why the same disease treated at two hospitals with two different products can produce two very different inpatient experiences without either being wrong. Other co-stimulatory domains — CD27, ICOS, OX40 — have been built into second-generation receptors, but their clinical progress has lagged the two validated platforms.
Generation Four: Putting the Cytokine Inside the Car
Armoured CAR-T cells, sometimes called TRUCKs, carry an inducible cytokine payload. The logic is specific to solid tumours: if the obstacle is a microenvironment that suppresses T cells, then a cell that delivers its own activating cytokine exactly where it engages antigen can overcome that suppression locally rather than systemically. IL-12, IL-15, IL-18 and IL-21 have all been tried in this role.
The design has produced real clinical signals. An IL-18-armoured CD19 construct, given to patients who had relapsed after earlier CD19 CAR-T, produced complete or partial responses in 81% at three months with a median response duration of 9.6 months. That was a 21-patient study, small enough that the number should be handled carefully, but it is the clearest demonstration so far that engineering can recover activity in a population that had already failed the same target.
The cost is the same logic running in reverse. A cell that secretes an inflammatory cytokine will sometimes secrete it in the wrong place or at the wrong magnitude, and the toxicity profile of an armoured product is not the toxicity profile of its unarmoured parent.
Generation Five: Precision Instead of Power
The current frontier uses synthetic biology to constrain rather than amplify. Logic gates restrict activation to cells carrying a defined combination of antigens, so that tissue expressing only one of them is spared. Affinity tuning lowers receptor sensitivity so that a high-density tumour antigen is engaged while a low-density normal antigen is not. Switchable and split receptors put antigen recognition and signalling under external control. Degron systems allow surface expression to be dialled down so that exhausted cells can rest.
The purpose of every one of these is to widen the therapeutic window — the gap between the dose that helps and the dose that harms. None of them is designed to produce a deeper response. That is worth stating plainly, because it reframes what these programmes are for: they are the technologies that make an aggressive target usable, not the technologies that make a modest target potent.
The Oldest Durability Data in the Field Is Not About Cancer
Long before CD19, the first human CAR-T trials targeted HIV. T cells were engineered with a CD4 extracellular domain or a gp41-specific scFv fused to CD3ζ, and a trial in HIV-infected adults demonstrated long-term persistence, homing to gut-associated tissue and local antiviral activity, without any change in systemic viral load.
In 2012, long-term follow-up of three of those trials reported that the engineered cells were still detectable more than ten years after infusion, with no evidence of vector-induced immortalisation or clonal expansion. That is the longest safety dataset in the field, and it predates every approved product. It answers a question that patients ask constantly and that no cancer trial has been running long enough to answer on its own: can these cells live in a person for a decade without becoming a second disease? In this cohort, yes.
What Solid Tumours Taught the Field
Solid tumours have resisted CAR-T far more stubbornly than blood cancers, and the recent signals have come less from better receptors than from better delivery. Four lessons recur across the trials that have worked.
Locoregional administration consistently produces stronger responses with less systemic toxicity than intravenous dosing, because the cells arrive where they are needed and the total body exposure falls. Repeated dosing helps, and so does cytokine support. Higher cell doses help. And tumour burden matters more than almost anything else: patients treated with low burden or earlier in their disease course respond better, which is the same lesson that the adjuvant gastric cancer data now being reported at conferences is built on.
Clinical benefit also correlates with how many CAR-T cells can be found inside the tumour, which means homing and local proliferation remain unsolved engineering problems rather than solved ones. Multi-antigen targeting and third-generation constructs with enhanced co-stimulation reduce escape, but responses in solid tumours are still frequently transient, and survival benefit is still limited.
Three Fronts That Cannot Be Skipped
The review closes on the three areas where the field's remaining work sits, and they are not equally mature.
On efficacy, the most productive recent interventions have been transcriptional. Forcing expression of FOXO1 holds cells in a less differentiated, memory-like state with better metabolic fitness under chronic stimulation. Overexpressing JUN counteracts the transcriptional remodelling that drives exhaustion. Deleting TOX, PRDM1 or NR4A1 produces cells resistant to terminal exhaustion. Metabolic reprogramming — through ACLY, PRODH2, PTPN2, the adenosine A1 receptor — improves mitochondrial fitness. And manufacturing conditions may matter as much as genetics: a product designed to preserve naive and stem-cell memory phenotypes produced clinical responses at one twenty-fifth of the conventional dose.
On safety, the tools are now credible. Drug-regulated switchable systems, chemically induced dimerisation, split receptors, degron-tagged receptors, synNotch circuits that restrict CAR expression to an antigen-defined microenvironment, and clinical-grade kill switches based on inducible caspase 9 are all in use or in trials. The trade-off is that a kill switch permanently removes the therapeutic cell, and RNA-encoded transient expression avoids that only by giving up persistence.
On access, two routes dominate. Allogeneic products from healthy donors attack cost and waiting time, at the price of graft-versus-host disease, host rejection, and the need for progressively more genomic edits — each of which raises its own concern about off-target effects and chromosomal rearrangements. And in vivo engineering, which delivers the CAR construct directly to a patient's own T cells, skips the laboratory entirely. That is the subject of a separate report on this site, and the first clinical series has now been published.
What to Do With All of This in a Consultation
The practical yield of a sixty-year history is small and specific. When you are offered a therapy, the generation number is not a question worth spending time on. The questions that change outcomes are narrower and more concrete, and they are the ones this site keeps returning to.
1. Which product, which target, and which co-stimulatory domain?
Not the generation — the specific construct. A CD28-based product and a 4-1BB-based product against the same antigen behave differently in the first two weeks, and that difference is worth knowing before you are admitted rather than after.
2. What is my tumour burden right now, and does the plan assume it stays there?
Burden is the most consistent predictor of response across the solid-tumour trials. If bridging therapy is available and not being used, ask why. If the answer is that waiting lists are long, that is a real answer — but it should be a stated one.
3. Is this an approved product or a trial, and what happens if it fails?
These carry different consent, different cost and different fallbacks. A trial that produces no response may also leave you ineligible for other trials for a period. Ask about the exit, not only the entrance.
4. What is the escalation plan for cytokine release, and who can authorise it at night?
Tocilizumab and steroids are standard; the variable is how fast they arrive. A centre that can answer this question with a protocol and a name is a different centre from one that answers it with reassurance.
5. If the receptor is fourth or fifth generation, what is being claimed — efficacy or safety?
Usually safety, and often correctly. Logic gates and switchable receptors exist to make hard targets survivable, not to make easy targets work. If the claim is efficacy, ask for the response rate in a comparable population.
A closing note on reading history as engineering.
The reason this history is worth a patient's attention is not nostalgia. It is that the field's habits are visible in it. The pattern of solving one failure and creating the next is not a sign that something is wrong; it is how a technology matures. What it does mean is that any single claim — a new generation, a new target, a new platform — should be read as a statement about one part of a system that has many parts. The parts that decide whether a particular person benefits are still the unglamorous ones: how much disease is present, how early the treatment arrives, whether the target sits on anything the body needs, and whether the team managing the first week has done it before.
Source: A guide to CAR T cell therapies: development, current status and future prospects. Nature Reviews Immunology. Published online July 2, 2026. The review's clinical data are drawn from trials published between 2006 and 2026; product approval counts in this article follow the review's own figures and may differ from national regulatory tallies, which this site tracks separately.